8016697: Use stubs to implement safefetch

Summary: Implement Safefetch as stub routines. This reduces compiler and os dependencies.
Reviewed-by: twisti, kvn

src/cpu/sparc/vm/stubGenerator_sparc.cpp

@@ -410,6 +410,51 @@ class StubGenerator: public StubCodeGenerator {
     return start;
   }
 
+  // Safefetch stubs.
+  void generate_safefetch(const char* name, int size, address* entry,
+                          address* fault_pc, address* continuation_pc) {
+    // safefetch signatures:
+    //   int      SafeFetch32(int*      adr, int      errValue);
+    //   intptr_t SafeFetchN (intptr_t* adr, intptr_t errValue);
+    //
+    // arguments:
+    //   o0 = adr
+    //   o1 = errValue
+    //
+    // result:
+    //   o0  = *adr or errValue
+
+    StubCodeMark mark(this, "StubRoutines", name);
+
+    // Entry point, pc or function descriptor.
+    __ align(CodeEntryAlignment);
+    *entry = __ pc();
+
+    __ mov(O0, G1);  // g1 = o0
+    __ mov(O1, O0);  // o0 = o1
+    // Load *adr into c_rarg1, may fault.
+    *fault_pc = __ pc();
+    switch (size) {
+      case 4:
+        // int32_t
+        __ ldsw(G1, 0, O0);  // o0 = [g1]
+        break;
+      case 8:
+        // int64_t
+        __ ldx(G1, 0, O0);   // o0 = [g1]
+        break;
+      default:
+        ShouldNotReachHere();
+    }
+
+    // return errValue or *adr
+    *continuation_pc = __ pc();
+    // By convention with the trap handler we ensure there is a non-CTI
+    // instruction in the trap shadow.
+    __ nop();
+    __ retl();
+    __ delayed()->nop();
+  }
 
   //------------------------------------------------------------------------------------------------------------------------
   // Continuation point for throwing of implicit exceptions that are not handled in
@@ -3315,6 +3360,14 @@ class StubGenerator: public StubCodeGenerator {
 
     // Don't initialize the platform math functions since sparc
     // doesn't have intrinsics for these operations.
+
+    // Safefetch stubs.
+    generate_safefetch("SafeFetch32", sizeof(int),     &StubRoutines::_safefetch32_entry,
+                                                       &StubRoutines::_safefetch32_fault_pc,
+                                                       &StubRoutines::_safefetch32_continuation_pc);
+    generate_safefetch("SafeFetchN", sizeof(intptr_t), &StubRoutines::_safefetchN_entry,
+                                                       &StubRoutines::_safefetchN_fault_pc,
+                                                       &StubRoutines::_safefetchN_continuation_pc);
   }
 
 

src/cpu/x86/vm/stubGenerator_x86_32.cpp

@@ -2766,6 +2766,39 @@ class StubGenerator: public StubCodeGenerator {
     return start;
   }
 
+  // Safefetch stubs.
+  void generate_safefetch(const char* name, int size, address* entry,
+                          address* fault_pc, address* continuation_pc) {
+    // safefetch signatures:
+    //   int      SafeFetch32(int*      adr, int      errValue);
+    //   intptr_t SafeFetchN (intptr_t* adr, intptr_t errValue);
+
+    StubCodeMark mark(this, "StubRoutines", name);
+
+    // Entry point, pc or function descriptor.
+    *entry = __ pc();
+
+    __ movl(rax, Address(rsp, 0x8));
+    __ movl(rcx, Address(rsp, 0x4));
+    // Load *adr into eax, may fault.
+    *fault_pc = __ pc();
+    switch (size) {
+      case 4:
+        // int32_t
+        __ movl(rax, Address(rcx, 0));
+        break;
+      case 8:
+        // int64_t
+        Unimplemented();
+        break;
+      default:
+        ShouldNotReachHere();
+    }
+
+    // Return errValue or *adr.
+    *continuation_pc = __ pc();
+    __ ret(0);
+  }
 
  public:
   // Information about frame layout at time of blocking runtime call.
@@ -2978,6 +3011,14 @@ class StubGenerator: public StubCodeGenerator {
       StubRoutines::_cipherBlockChaining_encryptAESCrypt = generate_cipherBlockChaining_encryptAESCrypt();
       StubRoutines::_cipherBlockChaining_decryptAESCrypt = generate_cipherBlockChaining_decryptAESCrypt();
     }
+
+    // Safefetch stubs.
+    generate_safefetch("SafeFetch32", sizeof(int), &StubRoutines::_safefetch32_entry,
+                                                   &StubRoutines::_safefetch32_fault_pc,
+                                                   &StubRoutines::_safefetch32_continuation_pc);
+    StubRoutines::_safefetchN_entry           = StubRoutines::_safefetch32_entry;
+    StubRoutines::_safefetchN_fault_pc        = StubRoutines::_safefetch32_fault_pc;
+    StubRoutines::_safefetchN_continuation_pc = StubRoutines::_safefetch32_continuation_pc;
   }
 
 

src/cpu/x86/vm/stubGenerator_x86_64.cpp

@@ -3357,7 +3357,45 @@ class StubGenerator: public StubCodeGenerator {
     return start;
   }
 
+  // Safefetch stubs.
+  void generate_safefetch(const char* name, int size, address* entry,
+                          address* fault_pc, address* continuation_pc) {
+    // safefetch signatures:
+    //   int      SafeFetch32(int*      adr, int      errValue);
+    //   intptr_t SafeFetchN (intptr_t* adr, intptr_t errValue);
+    //
+    // arguments:
+    //   c_rarg0 = adr
+    //   c_rarg1 = errValue
+    //
+    // result:
+    //   PPC_RET  = *adr or errValue
+
+    StubCodeMark mark(this, "StubRoutines", name);
+
+    // Entry point, pc or function descriptor.
+    *entry = __ pc();
+
+    // Load *adr into c_rarg1, may fault.
+    *fault_pc = __ pc();
+    switch (size) {
+      case 4:
+        // int32_t
+        __ movl(c_rarg1, Address(c_rarg0, 0));
+        break;
+      case 8:
+        // int64_t
+        __ movq(c_rarg1, Address(c_rarg0, 0));
+        break;
+      default:
+        ShouldNotReachHere();
+    }
 
+    // return errValue or *adr
+    *continuation_pc = __ pc();
+    __ movq(rax, c_rarg1);
+    __ ret(0);
+  }
 
   // This is a version of CBC/AES Decrypt which does 4 blocks in a loop at a time
   // to hide instruction latency
@@ -3833,6 +3871,14 @@ class StubGenerator: public StubCodeGenerator {
       StubRoutines::_cipherBlockChaining_encryptAESCrypt = generate_cipherBlockChaining_encryptAESCrypt();
       StubRoutines::_cipherBlockChaining_decryptAESCrypt = generate_cipherBlockChaining_decryptAESCrypt_Parallel();
     }
+
+    // Safefetch stubs.
+    generate_safefetch("SafeFetch32", sizeof(int),     &StubRoutines::_safefetch32_entry,
+                                                       &StubRoutines::_safefetch32_fault_pc,
+                                                       &StubRoutines::_safefetch32_continuation_pc);
+    generate_safefetch("SafeFetchN", sizeof(intptr_t), &StubRoutines::_safefetchN_entry,
+                                                       &StubRoutines::_safefetchN_fault_pc,
+                                                       &StubRoutines::_safefetchN_continuation_pc);
   }
 
  public:

src/os/windows/vm/os_windows.cpp

@@ -2317,6 +2317,11 @@ LONG WINAPI topLevelExceptionFilter(struct _EXCEPTION_POINTERS* exceptionInfo) {
 #endif
   Thread* t = ThreadLocalStorage::get_thread_slow();          // slow & steady
 
+  // Handle SafeFetch32 and SafeFetchN exceptions.
+  if (StubRoutines::is_safefetch_fault(pc)) {
+    return Handle_Exception(exceptionInfo, StubRoutines::continuation_for_safefetch_fault(pc));
+  }
+
 #ifndef _WIN64
   // Execution protection violation - win32 running on AMD64 only
   // Handled first to avoid misdiagnosis as a "normal" access violation;

src/os_cpu/bsd_x86/vm/bsd_x86_32.s

@@ -63,24 +63,6 @@ SYMBOL(fixcw):
         popl     %eax
         ret
 
-        .globl  SYMBOL(SafeFetch32), SYMBOL(Fetch32PFI), SYMBOL(Fetch32Resume)
-        .globl  SYMBOL(SafeFetchN)
-        ## TODO: avoid exposing Fetch32PFI and Fetch32Resume.
-        ## Instead, the signal handler would call a new SafeFetchTriage(FaultingEIP)
-        ## routine to vet the address.  If the address is the faulting LD then
-        ## SafeFetchTriage() would return the resume-at EIP, otherwise null.
-        ELF_TYPE(SafeFetch32,@function)
-        .p2align 4,,15
-SYMBOL(SafeFetch32):
-SYMBOL(SafeFetchN):
-         movl    0x8(%esp), %eax
-         movl    0x4(%esp), %ecx
-SYMBOL(Fetch32PFI):
-         movl    (%ecx), %eax
-SYMBOL(Fetch32Resume):
-         ret
-
-
         .globl  SYMBOL(SpinPause)
         ELF_TYPE(SpinPause,@function)
         .p2align 4,,15

src/os_cpu/bsd_x86/vm/bsd_x86_64.s

@@ -46,28 +46,6 @@
 
 	.text
 
-        .globl SYMBOL(SafeFetch32), SYMBOL(Fetch32PFI), SYMBOL(Fetch32Resume)
-        .p2align 4,,15
-        ELF_TYPE(SafeFetch32,@function)
-        // Prototype: int SafeFetch32 (int * Adr, int ErrValue) 
-SYMBOL(SafeFetch32):
-        movl    %esi, %eax
-SYMBOL(Fetch32PFI):
-        movl    (%rdi), %eax
-SYMBOL(Fetch32Resume):
-        ret
-
-        .globl SYMBOL(SafeFetchN), SYMBOL(FetchNPFI), SYMBOL(FetchNResume)
-        .p2align 4,,15
-        ELF_TYPE(SafeFetchN,@function)
-        // Prototype: intptr_t SafeFetchN (intptr_t * Adr, intptr_t ErrValue) 
-SYMBOL(SafeFetchN):
-        movq    %rsi, %rax
-SYMBOL(FetchNPFI):
-        movq    (%rdi), %rax
-SYMBOL(FetchNResume):
-        ret
-
         .globl SYMBOL(SpinPause)
         .p2align 4,,15
         ELF_TYPE(SpinPause,@function)

src/os_cpu/bsd_x86/vm/os_bsd_x86.cpp

@@ -385,13 +385,6 @@ enum {
   trap_page_fault = 0xE
 };
 
-extern "C" void Fetch32PFI () ;
-extern "C" void Fetch32Resume () ;
-#ifdef AMD64
-extern "C" void FetchNPFI () ;
-extern "C" void FetchNResume () ;
-#endif // AMD64
-
 extern "C" JNIEXPORT int
 JVM_handle_bsd_signal(int sig,
                         siginfo_t* info,
@@ -454,16 +447,10 @@ JVM_handle_bsd_signal(int sig,
   if (info != NULL && uc != NULL && thread != NULL) {
     pc = (address) os::Bsd::ucontext_get_pc(uc);
 
-    if (pc == (address) Fetch32PFI) {
-       uc->context_pc = intptr_t(Fetch32Resume) ;
-       return 1 ;
+    if (StubRoutines::is_safefetch_fault(pc)) {
+      uc->context_pc = intptr_t(StubRoutines::continuation_for_safefetch_fault(pc));
+      return 1;
     }
-#ifdef AMD64
-    if (pc == (address) FetchNPFI) {
-       uc->context_pc = intptr_t (FetchNResume) ;
-       return 1 ;
-    }
-#endif // AMD64
 
     // Handle ALL stack overflow variations here
     if (sig == SIGSEGV || sig == SIGBUS) {

src/os_cpu/linux_sparc/vm/linux_sparc.s

@@ -21,42 +21,6 @@
 # questions.
 #
 
-    # Prototype: int SafeFetch32 (int * adr, int ErrValue)
-    # The "ld" at Fetch32 is potentially faulting instruction.
-    # If the instruction traps the trap handler will arrange
-    # for control to resume at Fetch32Resume.  
-    # By convention with the trap handler we ensure there is a non-CTI
-    # instruction in the trap shadow.  
-        
-
-    .globl  SafeFetch32, Fetch32PFI, Fetch32Resume
-    .globl  SafeFetchN
-    .align  32
-    .type    SafeFetch32,@function
-SafeFetch32:        
-    mov     %o0, %g1
-    mov     %o1, %o0
-Fetch32PFI:
-    # <-- Potentially faulting instruction
-    ld      [%g1], %o0         
-Fetch32Resume:
-    nop
-    retl
-    nop
-
-    .globl  SafeFetchN, FetchNPFI, FetchNResume
-    .type    SafeFetchN,@function
-    .align  32
-SafeFetchN:
-    mov     %o0, %g1
-    mov     %o1, %o0
-FetchNPFI:
-    ldn     [%g1], %o0
-FetchNResume:
-    nop
-    retl
-    nop
-
     # Possibilities:
     # -- membar
     # -- CAS (SP + BIAS, G0, G0)

src/os_cpu/linux_sparc/vm/os_linux_sparc.cpp

@@ -366,18 +366,9 @@ intptr_t* os::Linux::ucontext_get_fp(ucontext_t *uc) {
 
 // Utility functions
 
-extern "C" void Fetch32PFI();
-extern "C" void Fetch32Resume();
-extern "C" void FetchNPFI();
-extern "C" void FetchNResume();
-
 inline static bool checkPrefetch(sigcontext* uc, address pc) {
-  if (pc == (address) Fetch32PFI) {
-    set_cont_address(uc, address(Fetch32Resume));
-    return true;
-  }
-  if (pc == (address) FetchNPFI) {
-    set_cont_address(uc, address(FetchNResume));
+  if (StubRoutines::is_safefetch_fault(pc)) {
+    set_cont_address(uc, address(StubRoutines::continuation_for_safefetch_fault(pc)));
     return true;
   }
   return false;

src/os_cpu/linux_x86/vm/linux_x86_32.s

@@ -42,24 +42,6 @@
 
 	.text
 
-        .globl  SafeFetch32, Fetch32PFI, Fetch32Resume
-        .globl  SafeFetchN
-        ## TODO: avoid exposing Fetch32PFI and Fetch32Resume.
-        ## Instead, the signal handler would call a new SafeFetchTriage(FaultingEIP)
-        ## routine to vet the address.  If the address is the faulting LD then
-        ## SafeFetchTriage() would return the resume-at EIP, otherwise null.
-	.type    SafeFetch32,@function
-        .p2align 4,,15
-SafeFetch32:
-SafeFetchN:
-         movl    0x8(%esp), %eax
-         movl    0x4(%esp), %ecx
-Fetch32PFI:
-         movl    (%ecx), %eax
-Fetch32Resume:
-         ret
-
-
         .globl  SpinPause
 	.type   SpinPause,@function
         .p2align 4,,15

src/os_cpu/linux_x86/vm/linux_x86_64.s

@@ -38,28 +38,6 @@
 
 	.text
 
-        .globl SafeFetch32, Fetch32PFI, Fetch32Resume
-        .align  16
-        .type   SafeFetch32,@function
-        // Prototype: int SafeFetch32 (int * Adr, int ErrValue) 
-SafeFetch32:
-        movl    %esi, %eax
-Fetch32PFI:
-        movl    (%rdi), %eax
-Fetch32Resume:
-        ret
-
-        .globl SafeFetchN, FetchNPFI, FetchNResume
-        .align  16
-        .type   SafeFetchN,@function
-        // Prototype: intptr_t SafeFetchN (intptr_t * Adr, intptr_t ErrValue) 
-SafeFetchN:
-        movq    %rsi, %rax
-FetchNPFI:
-        movq    (%rdi), %rax
-FetchNResume:
-        ret
-
         .globl SpinPause
         .align 16
         .type  SpinPause,@function

src/os_cpu/linux_x86/vm/os_linux_x86.cpp

@@ -209,13 +209,6 @@ enum {
   trap_page_fault = 0xE
 };
 
-extern "C" void Fetch32PFI () ;
-extern "C" void Fetch32Resume () ;
-#ifdef AMD64
-extern "C" void FetchNPFI () ;
-extern "C" void FetchNResume () ;
-#endif // AMD64
-
 extern "C" JNIEXPORT int
 JVM_handle_linux_signal(int sig,
                         siginfo_t* info,
@@ -278,16 +271,10 @@ JVM_handle_linux_signal(int sig,
   if (info != NULL && uc != NULL && thread != NULL) {
     pc = (address) os::Linux::ucontext_get_pc(uc);
 
-    if (pc == (address) Fetch32PFI) {
-       uc->uc_mcontext.gregs[REG_PC] = intptr_t(Fetch32Resume) ;
-       return 1 ;
+    if (StubRoutines::is_safefetch_fault(pc)) {
+      uc->uc_mcontext.gregs[REG_PC] = intptr_t(StubRoutines::continuation_for_safefetch_fault(pc));
+      return 1;
     }
-#ifdef AMD64
-    if (pc == (address) FetchNPFI) {
-       uc->uc_mcontext.gregs[REG_PC] = intptr_t (FetchNResume) ;
-       return 1 ;
-    }
-#endif // AMD64
 
 #ifndef AMD64
     // Halt if SI_KERNEL before more crashes get misdiagnosed as Java bugs

src/os_cpu/solaris_sparc/vm/os_solaris_sparc.cpp

@@ -303,11 +303,6 @@ bool os::is_allocatable(size_t bytes) {
 #endif
 }
 
-extern "C" void Fetch32PFI () ;
-extern "C" void Fetch32Resume () ;
-extern "C" void FetchNPFI () ;
-extern "C" void FetchNResume () ;
-
 extern "C" JNIEXPORT int
 JVM_handle_solaris_signal(int sig, siginfo_t* info, void* ucVoid,
                           int abort_if_unrecognized) {
@@ -379,17 +374,10 @@ JVM_handle_solaris_signal(int sig, siginfo_t* info, void* ucVoid,
     npc = (address) uc->uc_mcontext.gregs[REG_nPC];
 
     // SafeFetch() support
-    // Implemented with either a fixed set of addresses such
-    // as Fetch32*, or with Thread._OnTrap.
-    if (uc->uc_mcontext.gregs[REG_PC] == intptr_t(Fetch32PFI)) {
-      uc->uc_mcontext.gregs [REG_PC]  = intptr_t(Fetch32Resume) ;
-      uc->uc_mcontext.gregs [REG_nPC] = intptr_t(Fetch32Resume) + 4 ;
-      return true ;
-    }
-    if (uc->uc_mcontext.gregs[REG_PC] == intptr_t(FetchNPFI)) {
-      uc->uc_mcontext.gregs [REG_PC]  = intptr_t(FetchNResume) ;
-      uc->uc_mcontext.gregs [REG_nPC] = intptr_t(FetchNResume) + 4 ;
-      return true ;
+    if (StubRoutines::is_safefetch_fault(pc)) {
+      uc->uc_mcontext.gregs[REG_PC] = intptr_t(StubRoutines::continuation_for_safefetch_fault(pc));
+      uc->uc_mcontext.gregs[REG_nPC] = uc->uc_mcontext.gregs[REG_PC] + 4;
+      return 1;
     }
 
     // Handle ALL stack overflow variations here

src/os_cpu/solaris_sparc/vm/solaris_sparc.s

@@ -21,47 +21,6 @@
 !! questions.
 !!
 
-    !! Prototype: int SafeFetch32 (int * adr, int ErrValue)
-    !! The "ld" at Fetch32 is potentially faulting instruction.
-    !! If the instruction traps the trap handler will arrange
-    !! for control to resume at Fetch32Resume.  
-    !! By convention with the trap handler we ensure there is a non-CTI
-    !! instruction in the trap shadow.  
-    !!
-    !! The reader might be tempted to move this service to .il.
-    !! Don't.  Sun's CC back-end reads and optimize code emitted 
-    !! by the .il "call", in some cases optimizing the code, completely eliding it,
-    !! or by moving the code from the "call site". 
-        
-     !! ASM better know we may use G6 for our own purposes
-    .register %g6, #ignore
-        
-    .globl  SafeFetch32
-    .align  32
-    .global Fetch32PFI, Fetch32Resume 
-SafeFetch32:
-    mov     %o0, %g1
-    mov     %o1, %o0
-Fetch32PFI:
-    ld      [%g1], %o0          !! <-- Potentially faulting instruction
-Fetch32Resume:
-    nop
-    retl
-    nop
-
-    .globl  SafeFetchN
-    .align  32
-    .globl  FetchNPFI, FetchNResume
-SafeFetchN:
-    mov     %o0, %g1
-    mov     %o1, %o0
-FetchNPFI:
-    ldn     [%g1], %o0
-FetchNResume:
-    nop
-    retl
-    nop
-
     !! Possibilities:
     !! -- membar
     !! -- CAS (SP + BIAS, G0, G0)

src/os_cpu/solaris_x86/vm/os_solaris_x86.cpp

@@ -352,13 +352,6 @@ bool os::is_allocatable(size_t bytes) {
 
 }
 
-extern "C" void Fetch32PFI () ;
-extern "C" void Fetch32Resume () ;
-#ifdef AMD64
-extern "C" void FetchNPFI () ;
-extern "C" void FetchNResume () ;
-#endif // AMD64
-
 extern "C" JNIEXPORT int
 JVM_handle_solaris_signal(int sig, siginfo_t* info, void* ucVoid,
                           int abort_if_unrecognized) {
@@ -436,17 +429,10 @@ JVM_handle_solaris_signal(int sig, siginfo_t* info, void* ucVoid,
     // factor me: getPCfromContext
     pc = (address) uc->uc_mcontext.gregs[REG_PC];
 
-    // SafeFetch32() support
-    if (pc == (address) Fetch32PFI) {
-      uc->uc_mcontext.gregs[REG_PC] = intptr_t(Fetch32Resume) ;
-      return true ;
-    }
-#ifdef AMD64
-    if (pc == (address) FetchNPFI) {
-       uc->uc_mcontext.gregs [REG_PC] = intptr_t(FetchNResume) ;
-       return true ;
+    if (StubRoutines::is_safefetch_fault(pc)) {
+      uc->uc_mcontext.gregs[REG_PC] = intptr_t(StubRoutines::continuation_for_safefetch_fault(pc));
+      return true;
     }
-#endif // AMD64
 
     // Handle ALL stack overflow variations here
     if (sig == SIGSEGV && info->si_code == SEGV_ACCERR) {

src/os_cpu/solaris_x86/vm/solaris_x86_32.s

@@ -54,20 +54,6 @@ fixcw:
 	popl     %eax
 	ret
 
-       .align  16
-       .globl  SafeFetch32
-       .globl  SafeFetchN
-       .globl  Fetch32PFI, Fetch32Resume
-SafeFetch32:
-SafeFetchN:
-        movl    0x8(%esp), %eax
-        movl    0x4(%esp), %ecx
-Fetch32PFI:
-        movl    (%ecx), %eax
-Fetch32Resume:
-        ret
-
-
         .align  16
         .globl  SpinPause
 SpinPause:

src/os_cpu/solaris_x86/vm/solaris_x86_64.s

@@ -21,54 +21,34 @@
 / questions.
 /
 
-	.globl fs_load
-	.globl fs_thread
+        .globl fs_load
+        .globl fs_thread
 
         // NOTE WELL!  The _Copy functions are called directly
-	// from server-compiler-generated code via CallLeafNoFP,
-	// which means that they *must* either not use floating
-	// point or use it in the same manner as does the server
-	// compiler.
+        // from server-compiler-generated code via CallLeafNoFP,
+        // which means that they *must* either not use floating
+        // point or use it in the same manner as does the server
+        // compiler.
 
         .globl _Copy_arrayof_conjoint_bytes
         .globl _Copy_conjoint_jshorts_atomic
-	.globl _Copy_arrayof_conjoint_jshorts
+        .globl _Copy_arrayof_conjoint_jshorts
         .globl _Copy_conjoint_jints_atomic
         .globl _Copy_arrayof_conjoint_jints
-	.globl _Copy_conjoint_jlongs_atomic
+        .globl _Copy_conjoint_jlongs_atomic
         .globl _Copy_arrayof_conjoint_jlongs
 
-	.section .text,"ax"
+        .section .text,"ax"
 
         / Fast thread accessors, used by threadLS_solaris_amd64.cpp
-	.align   16
+        .align   16
 fs_load:
-	movq %fs:(%rdi),%rax
-	ret
-
-	.align   16
-fs_thread:
-	movq %fs:0x0,%rax
-	ret
-
-        .globl SafeFetch32, Fetch32PFI, Fetch32Resume
-        .align  16
-        // Prototype: int SafeFetch32 (int * Adr, int ErrValue) 
-SafeFetch32:
-        movl    %esi, %eax
-Fetch32PFI:
-        movl    (%rdi), %eax
-Fetch32Resume:
+        movq %fs:(%rdi),%rax
         ret
 
-        .globl SafeFetchN, FetchNPFI, FetchNResume
-        .align  16
-        // Prototype: intptr_t SafeFetchN (intptr_t * Adr, intptr_t ErrValue) 
-SafeFetchN:
-        movq    %rsi, %rax
-FetchNPFI:
-        movq    (%rdi), %rax
-FetchNResume:
+        .align   16
+fs_thread:
+        movq %fs:0x0,%rax
         ret
 
         .globl  SpinPause
@@ -78,7 +58,7 @@ SpinPause:
         nop
         movq    $1, %rax
         ret
-        
+
 
         / Support for void Copy::arrayof_conjoint_bytes(void* from,
         /                                               void* to,
@@ -340,7 +320,7 @@ aci_CopyLeft:
         addq     $4,%rdx
         jg       1b
         ret
-	
+
         / Support for void Copy::arrayof_conjoint_jlongs(jlong* from,
         /                                                jlong* to,
         /                                                size_t count)

src/os_cpu/windows_x86/vm/os_windows_x86.cpp

@@ -518,24 +518,6 @@ void os::print_register_info(outputStream *st, void *context) {
   st->cr();
 }
 
-extern "C" int SafeFetch32 (int * adr, int Err) {
-   int rv = Err ;
-   _try {
-       rv = *((volatile int *) adr) ;
-   } __except(EXCEPTION_EXECUTE_HANDLER) {
-   }
-   return rv ;
-}
-
-extern "C" intptr_t SafeFetchN (intptr_t * adr, intptr_t Err) {
-   intptr_t rv = Err ;
-   _try {
-       rv = *((volatile intptr_t *) adr) ;
-   } __except(EXCEPTION_EXECUTE_HANDLER) {
-   }
-   return rv ;
-}
-
 extern "C" int SpinPause () {
 #ifdef AMD64
    return 0 ;

src/share/vm/runtime/os.hpp

@@ -915,8 +915,6 @@ class os: AllStatic {
 // of the global SpinPause() with C linkage.
 // It'd also be eligible for inlining on many platforms.
 
-extern "C" int SpinPause () ;
-extern "C" int SafeFetch32 (int * adr, int errValue) ;
-extern "C" intptr_t SafeFetchN (intptr_t * adr, intptr_t errValue) ;
+extern "C" int SpinPause();
 
 #endif // SHARE_VM_RUNTIME_OS_HPP

src/share/vm/runtime/stubRoutines.cpp

@@ -136,6 +136,13 @@ double (* StubRoutines::_intrinsic_sin   )(double) = NULL;
 double (* StubRoutines::_intrinsic_cos   )(double) = NULL;
 double (* StubRoutines::_intrinsic_tan   )(double) = NULL;
 
+address StubRoutines::_safefetch32_entry                 = NULL;
+address StubRoutines::_safefetch32_fault_pc              = NULL;
+address StubRoutines::_safefetch32_continuation_pc       = NULL;
+address StubRoutines::_safefetchN_entry                  = NULL;
+address StubRoutines::_safefetchN_fault_pc               = NULL;
+address StubRoutines::_safefetchN_continuation_pc        = NULL;
+
 // Initialization
 //
 // Note: to break cycle with universe initialization, stubs are generated in two phases.

src/share/vm/runtime/stubRoutines.hpp

@@ -221,6 +221,14 @@ class StubRoutines: AllStatic {
   static double (*_intrinsic_cos)(double);
   static double (*_intrinsic_tan)(double);
 
+  // Safefetch stubs.
+  static address _safefetch32_entry;
+  static address _safefetch32_fault_pc;
+  static address _safefetch32_continuation_pc;
+  static address _safefetchN_entry;
+  static address _safefetchN_fault_pc;
+  static address _safefetchN_continuation_pc;
+
  public:
   // Initialization/Testing
   static void    initialize1();                            // must happen before universe::genesis
@@ -381,6 +389,34 @@ class StubRoutines: AllStatic {
     return _intrinsic_tan(d);
   }
 
+  //
+  // Safefetch stub support
+  //
+
+  typedef int      (*SafeFetch32Stub)(int*      adr, int      errValue);
+  typedef intptr_t (*SafeFetchNStub) (intptr_t* adr, intptr_t errValue);
+
+  static SafeFetch32Stub SafeFetch32_stub() { return CAST_TO_FN_PTR(SafeFetch32Stub, _safefetch32_entry); }
+  static SafeFetchNStub  SafeFetchN_stub()  { return CAST_TO_FN_PTR(SafeFetchNStub,  _safefetchN_entry); }
+
+  static bool is_safefetch_fault(address pc) {
+    return pc != NULL &&
+          (pc == _safefetch32_fault_pc ||
+           pc == _safefetchN_fault_pc);
+  }
+
+  static address continuation_for_safefetch_fault(address pc) {
+    assert(_safefetch32_continuation_pc != NULL &&
+           _safefetchN_continuation_pc  != NULL,
+           "not initialized");
+
+    if (pc == _safefetch32_fault_pc) return _safefetch32_continuation_pc;
+    if (pc == _safefetchN_fault_pc)  return _safefetchN_continuation_pc;
+
+    ShouldNotReachHere();
+    return NULL;
+  }
+
   //
   // Default versions of the above arraycopy functions for platforms which do
   // not have specialized versions
@@ -400,4 +436,15 @@ class StubRoutines: AllStatic {
   static void arrayof_oop_copy_uninit(HeapWord* src, HeapWord* dest, size_t count);
 };
 
+// Safefetch allows to load a value from a location that's not known
+// to be valid. If the load causes a fault, the error value is returned.
+inline int SafeFetch32(int* adr, int errValue) {
+  assert(StubRoutines::SafeFetch32_stub(), "stub not yet generated");
+  return StubRoutines::SafeFetch32_stub()(adr, errValue);
+}
+inline intptr_t SafeFetchN(intptr_t* adr, intptr_t errValue) {
+  assert(StubRoutines::SafeFetchN_stub(), "stub not yet generated");
+  return StubRoutines::SafeFetchN_stub()(adr, errValue);
+}
+
 #endif // SHARE_VM_RUNTIME_STUBROUTINES_HPP