Mercurial > jdk8-mips64-public > hotspot / changeset
changeset
8068945: Use RBP register as proper frame pointer in JIT compiled code on x86
Fri, 29 May 2015 10:58:45 +0200
- author
- zmajo
- date
- Fri, 29 May 2015 10:58:45 +0200
- changeset 7854
- e8260b6328fb
- parent 7853
- a1642365d69f
- child 7855
- 62df92c92d33
8068945: Use RBP register as proper frame pointer in JIT compiled code on x86
Summary: Introduce the PreserveFramePointer flag to control if RBP is used as the frame pointer or as a general purpose register.
Reviewed-by: kvn, roland, dlong, enevill, shade
1.1 --- a/agent/src/share/classes/sun/jvm/hotspot/runtime/x86/X86Frame.java Fri Mar 27 10:57:42 2015 +0100 1.2 +++ b/agent/src/share/classes/sun/jvm/hotspot/runtime/x86/X86Frame.java Fri May 29 10:58:45 2015 +0200 1.3 @@ -317,26 +317,17 @@ 1.4 //------------------------------------------------------------------------------ 1.5 // frame::adjust_unextended_sp 1.6 private void adjustUnextendedSP() { 1.7 - // If we are returning to a compiled MethodHandle call site, the 1.8 - // saved_fp will in fact be a saved value of the unextended SP. The 1.9 - // simplest way to tell whether we are returning to such a call site 1.10 - // is as follows: 1.11 + // On x86, sites calling method handle intrinsics and lambda forms are treated 1.12 + // as any other call site. Therefore, no special action is needed when we are 1.13 + // returning to any of these call sites. 1.14 1.15 CodeBlob cb = cb(); 1.16 NMethod senderNm = (cb == null) ? null : cb.asNMethodOrNull(); 1.17 if (senderNm != null) { 1.18 - // If the sender PC is a deoptimization point, get the original 1.19 - // PC. For MethodHandle call site the unextended_sp is stored in 1.20 - // saved_fp. 1.21 - if (senderNm.isDeoptMhEntry(getPC())) { 1.22 - // DEBUG_ONLY(verifyDeoptMhOriginalPc(senderNm, getFP())); 1.23 - raw_unextendedSP = getFP(); 1.24 - } 1.25 - else if (senderNm.isDeoptEntry(getPC())) { 1.26 - // DEBUG_ONLY(verifyDeoptOriginalPc(senderNm, raw_unextendedSp)); 1.27 - } 1.28 - else if (senderNm.isMethodHandleReturn(getPC())) { 1.29 - raw_unextendedSP = getFP(); 1.30 + // If the sender PC is a deoptimization point, get the original PC. 1.31 + if (senderNm.isDeoptEntry(getPC()) || 1.32 + senderNm.isDeoptMhEntry(getPC())) { 1.33 + // DEBUG_ONLY(verifyDeoptriginalPc(senderNm, raw_unextendedSp)); 1.34 } 1.35 } 1.36 }
2.1 --- a/src/cpu/ppc/vm/globals_ppc.hpp Fri Mar 27 10:57:42 2015 +0100 2.2 +++ b/src/cpu/ppc/vm/globals_ppc.hpp Fri May 29 10:58:45 2015 +0200 2.3 @@ -55,6 +55,8 @@ 2.4 2.5 define_pd_global(bool, UseMembar, false); 2.6 2.7 +define_pd_global(bool, PreserveFramePointer, false); 2.8 + 2.9 // GC Ergo Flags 2.10 define_pd_global(uintx, CMSYoungGenPerWorker, 16*M); // Default max size of CMS young gen, per GC worker thread. 2.11
3.1 --- a/src/cpu/sparc/vm/globals_sparc.hpp Fri Mar 27 10:57:42 2015 +0100 3.2 +++ b/src/cpu/sparc/vm/globals_sparc.hpp Fri May 29 10:58:45 2015 +0200 3.3 @@ -74,6 +74,8 @@ 3.4 3.5 define_pd_global(bool, UseMembar, false); 3.6 3.7 +define_pd_global(bool, PreserveFramePointer, false); 3.8 + 3.9 // GC Ergo Flags 3.10 define_pd_global(uintx, CMSYoungGenPerWorker, 16*M); // default max size of CMS young gen, per GC worker thread 3.11
4.1 --- a/src/cpu/x86/vm/assembler_x86.hpp Fri Mar 27 10:57:42 2015 +0100 4.2 +++ b/src/cpu/x86/vm/assembler_x86.hpp Fri May 29 10:58:45 2015 +0200 4.3 @@ -141,8 +141,10 @@ 4.4 4.5 #endif // _LP64 4.6 4.7 -// JSR 292 fixed register usages: 4.8 -REGISTER_DECLARATION(Register, rbp_mh_SP_save, rbp); 4.9 +// JSR 292 4.10 +// On x86, the SP does not have to be saved when invoking method handle intrinsics 4.11 +// or compiled lambda forms. We indicate that by setting rbp_mh_SP_save to noreg. 4.12 +REGISTER_DECLARATION(Register, rbp_mh_SP_save, noreg); 4.13 4.14 // Address is an abstraction used to represent a memory location 4.15 // using any of the amd64 addressing modes with one object.
5.1 --- a/src/cpu/x86/vm/c1_FrameMap_x86.cpp Fri Mar 27 10:57:42 2015 +0100 5.2 +++ b/src/cpu/x86/vm/c1_FrameMap_x86.cpp Fri May 29 10:58:45 2015 +0200 5.3 @@ -343,14 +343,13 @@ 5.4 return FrameMap::rsp_opr; 5.5 } 5.6 5.7 - 5.8 // JSR 292 5.9 +// On x86, there is no need to save the SP, because neither 5.10 +// method handle intrinsics, nor compiled lambda forms modify it. 5.11 LIR_Opr FrameMap::method_handle_invoke_SP_save_opr() { 5.12 - assert(rbp == rbp_mh_SP_save, "must be same register"); 5.13 - return rbp_opr; 5.14 + return LIR_OprFact::illegalOpr; 5.15 } 5.16 5.17 - 5.18 bool FrameMap::validate_frame() { 5.19 return true; 5.20 }
6.1 --- a/src/cpu/x86/vm/c1_MacroAssembler_x86.cpp Fri Mar 27 10:57:42 2015 +0100 6.2 +++ b/src/cpu/x86/vm/c1_MacroAssembler_x86.cpp Fri May 29 10:58:45 2015 +0200 6.3 @@ -359,6 +359,9 @@ 6.4 generate_stack_overflow_check(bang_size_in_bytes); 6.5 6.6 push(rbp); 6.7 + if (PreserveFramePointer) { 6.8 + mov(rbp, rsp); 6.9 + } 6.10 #ifdef TIERED 6.11 // c2 leaves fpu stack dirty. Clean it on entry 6.12 if (UseSSE < 2 ) {
7.1 --- a/src/cpu/x86/vm/c1_Runtime1_x86.cpp Fri Mar 27 10:57:42 2015 +0100 7.2 +++ b/src/cpu/x86/vm/c1_Runtime1_x86.cpp Fri May 29 10:58:45 2015 +0200 7.3 @@ -754,14 +754,9 @@ 7.4 // WIN64_ONLY: No need to add frame::arg_reg_save_area_bytes to SP 7.5 // since we do a leave anyway. 7.6 7.7 - // Pop the return address since we are possibly changing SP (restoring from BP). 7.8 + // Pop the return address. 7.9 __ leave(); 7.10 __ pop(rcx); 7.11 - 7.12 - // Restore SP from BP if the exception PC is a method handle call site. 7.13 - NOT_LP64(__ get_thread(thread);) 7.14 - __ cmpl(Address(thread, JavaThread::is_method_handle_return_offset()), 0); 7.15 - __ cmovptr(Assembler::notEqual, rsp, rbp_mh_SP_save); 7.16 __ jmp(rcx); // jump to exception handler 7.17 break; 7.18 default: ShouldNotReachHere(); 7.19 @@ -832,11 +827,6 @@ 7.20 // the pop is also necessary to simulate the effect of a ret(0) 7.21 __ pop(exception_pc); 7.22 7.23 - // Restore SP from BP if the exception PC is a method handle call site. 7.24 - NOT_LP64(__ get_thread(thread);) 7.25 - __ cmpl(Address(thread, JavaThread::is_method_handle_return_offset()), 0); 7.26 - __ cmovptr(Assembler::notEqual, rsp, rbp_mh_SP_save); 7.27 - 7.28 // continue at exception handler (return address removed) 7.29 // note: do *not* remove arguments when unwinding the 7.30 // activation since the caller assumes having
8.1 --- a/src/cpu/x86/vm/frame_x86.cpp Fri Mar 27 10:57:42 2015 +0100 8.2 +++ b/src/cpu/x86/vm/frame_x86.cpp Fri May 29 10:58:45 2015 +0200 8.3 @@ -216,7 +216,8 @@ 8.4 if (sender_blob->is_nmethod()) { 8.5 nmethod* nm = sender_blob->as_nmethod_or_null(); 8.6 if (nm != NULL) { 8.7 - if (nm->is_deopt_mh_entry(sender_pc) || nm->is_deopt_entry(sender_pc)) { 8.8 + if (nm->is_deopt_mh_entry(sender_pc) || nm->is_deopt_entry(sender_pc) || 8.9 + nm->method()->is_method_handle_intrinsic()) { 8.10 return false; 8.11 } 8.12 } 8.13 @@ -383,10 +384,9 @@ 8.14 // frame::verify_deopt_original_pc 8.15 // 8.16 // Verifies the calculated original PC of a deoptimization PC for the 8.17 -// given unextended SP. The unextended SP might also be the saved SP 8.18 -// for MethodHandle call sites. 8.19 +// given unextended SP. 8.20 #ifdef ASSERT 8.21 -void frame::verify_deopt_original_pc(nmethod* nm, intptr_t* unextended_sp, bool is_method_handle_return) { 8.22 +void frame::verify_deopt_original_pc(nmethod* nm, intptr_t* unextended_sp) { 8.23 frame fr; 8.24 8.25 // This is ugly but it's better than to change {get,set}_original_pc 8.26 @@ -396,33 +396,23 @@ 8.27 8.28 address original_pc = nm->get_original_pc(&fr); 8.29 assert(nm->insts_contains(original_pc), "original PC must be in nmethod"); 8.30 - assert(nm->is_method_handle_return(original_pc) == is_method_handle_return, "must be"); 8.31 } 8.32 #endif 8.33 8.34 //------------------------------------------------------------------------------ 8.35 // frame::adjust_unextended_sp 8.36 void frame::adjust_unextended_sp() { 8.37 - // If we are returning to a compiled MethodHandle call site, the 8.38 - // saved_fp will in fact be a saved value of the unextended SP. The 8.39 - // simplest way to tell whether we are returning to such a call site 8.40 - // is as follows: 8.41 + // On x86, sites calling method handle intrinsics and lambda forms are treated 8.42 + // as any other call site. Therefore, no special action is needed when we are 8.43 + // returning to any of these call sites. 8.44 8.45 nmethod* sender_nm = (_cb == NULL) ? NULL : _cb->as_nmethod_or_null(); 8.46 if (sender_nm != NULL) { 8.47 - // If the sender PC is a deoptimization point, get the original 8.48 - // PC. For MethodHandle call site the unextended_sp is stored in 8.49 - // saved_fp. 8.50 - if (sender_nm->is_deopt_mh_entry(_pc)) { 8.51 - DEBUG_ONLY(verify_deopt_mh_original_pc(sender_nm, _fp)); 8.52 - _unextended_sp = _fp; 8.53 - } 8.54 - else if (sender_nm->is_deopt_entry(_pc)) { 8.55 + // If the sender PC is a deoptimization point, get the original PC. 8.56 + if (sender_nm->is_deopt_entry(_pc) || 8.57 + sender_nm->is_deopt_mh_entry(_pc)) { 8.58 DEBUG_ONLY(verify_deopt_original_pc(sender_nm, _unextended_sp)); 8.59 } 8.60 - else if (sender_nm->is_method_handle_return(_pc)) { 8.61 - _unextended_sp = _fp; 8.62 - } 8.63 } 8.64 } 8.65
9.1 --- a/src/cpu/x86/vm/frame_x86.hpp Fri Mar 27 10:57:42 2015 +0100 9.2 +++ b/src/cpu/x86/vm/frame_x86.hpp Fri May 29 10:58:45 2015 +0200 9.3 @@ -76,11 +76,11 @@ 9.4 // [locals and parameters ] 9.5 // <- sender sp 9.6 9.7 -// [1] When the c++ interpreter calls a new method it returns to the frame 9.8 +// [1] When the C++ interpreter calls a new method it returns to the frame 9.9 // manager which allocates a new frame on the stack. In that case there 9.10 // is no real callee of this newly allocated frame. The frame manager is 9.11 -// aware of the additional frame(s) and will pop them as nested calls 9.12 -// complete. Howevers tTo make it look good in the debugger the frame 9.13 +// aware of the additional frame(s) and will pop them as nested calls 9.14 +// complete. However, to make it look good in the debugger the frame 9.15 // manager actually installs a dummy pc pointing to RecursiveInterpreterActivation 9.16 // with a fake interpreter_state* parameter to make it easy to debug 9.17 // nested calls. 9.18 @@ -88,7 +88,7 @@ 9.19 // Note that contrary to the layout for the assembly interpreter the 9.20 // expression stack allocated for the C++ interpreter is full sized. 9.21 // However this is not as bad as it seems as the interpreter frame_manager 9.22 -// will truncate the unused space on succesive method calls. 9.23 +// will truncate the unused space on successive method calls. 9.24 // 9.25 // ------------------------------ C++ interpreter ---------------------------------------- 9.26 9.27 @@ -172,10 +172,7 @@ 9.28 9.29 #ifdef ASSERT 9.30 // Used in frame::sender_for_{interpreter,compiled}_frame 9.31 - static void verify_deopt_original_pc( nmethod* nm, intptr_t* unextended_sp, bool is_method_handle_return = false); 9.32 - static void verify_deopt_mh_original_pc(nmethod* nm, intptr_t* unextended_sp) { 9.33 - verify_deopt_original_pc(nm, unextended_sp, true); 9.34 - } 9.35 + static void verify_deopt_original_pc(nmethod* nm, intptr_t* unextended_sp); 9.36 #endif 9.37 9.38 public:
10.1 --- a/src/cpu/x86/vm/frame_x86.inline.hpp Fri Mar 27 10:57:42 2015 +0100 10.2 +++ b/src/cpu/x86/vm/frame_x86.inline.hpp Fri May 29 10:58:45 2015 +0200 10.3 @@ -93,7 +93,7 @@ 10.4 // find_blob call. This is also why we can have no asserts on the validity 10.5 // of the pc we find here. AsyncGetCallTrace -> pd_get_top_frame_for_signal_handler 10.6 // -> pd_last_frame should use a specialized version of pd_last_frame which could 10.7 - // call a specilaized frame constructor instead of this one. 10.8 + // call a specialized frame constructor instead of this one. 10.9 // Then we could use the assert below. However this assert is of somewhat dubious 10.10 // value. 10.11 // assert(_pc != NULL, "no pc?");
11.1 --- a/src/cpu/x86/vm/globals_x86.hpp Fri Mar 27 10:57:42 2015 +0100 11.2 +++ b/src/cpu/x86/vm/globals_x86.hpp Fri May 29 10:58:45 2015 +0200 11.3 @@ -82,6 +82,8 @@ 11.4 11.5 define_pd_global(uintx, TypeProfileLevel, 111); 11.6 11.7 +define_pd_global(bool, PreserveFramePointer, false); 11.8 + 11.9 #define ARCH_FLAGS(develop, product, diagnostic, experimental, notproduct) \ 11.10 \ 11.11 develop(bool, IEEEPrecision, true, \
12.1 --- a/src/cpu/x86/vm/macroAssembler_x86.cpp Fri Mar 27 10:57:42 2015 +0100 12.2 +++ b/src/cpu/x86/vm/macroAssembler_x86.cpp Fri May 29 10:58:45 2015 +0200 12.3 @@ -6122,6 +6122,10 @@ 12.4 // We always push rbp, so that on return to interpreter rbp, will be 12.5 // restored correctly and we can correct the stack. 12.6 push(rbp); 12.7 + // Save caller's stack pointer into RBP if the frame pointer is preserved. 12.8 + if (PreserveFramePointer) { 12.9 + mov(rbp, rsp); 12.10 + } 12.11 // Remove word for ebp 12.12 framesize -= wordSize; 12.13 12.14 @@ -6136,6 +6140,11 @@ 12.15 // Save RBP register now. 12.16 framesize -= wordSize; 12.17 movptr(Address(rsp, framesize), rbp); 12.18 + // Save caller's stack pointer into RBP if the frame pointer is preserved. 12.19 + if (PreserveFramePointer) { 12.20 + movptr(rbp, rsp); 12.21 + addptr(rbp, framesize + wordSize); 12.22 + } 12.23 } 12.24 12.25 if (VerifyStackAtCalls) { // Majik cookie to verify stack depth
13.1 --- a/src/cpu/x86/vm/methodHandles_x86.cpp Fri Mar 27 10:57:42 2015 +0100 13.2 +++ b/src/cpu/x86/vm/methodHandles_x86.cpp Fri May 29 10:58:45 2015 +0200 13.3 @@ -373,7 +373,7 @@ 13.4 // member_reg - MemberName that was the trailing argument 13.5 // temp1_recv_klass - klass of stacked receiver, if needed 13.6 // rsi/r13 - interpreter linkage (if interpreted) 13.7 - // rcx, rdx, rsi, rdi, r8, r8 - compiler arguments (if compiled) 13.8 + // rcx, rdx, rsi, rdi, r8 - compiler arguments (if compiled) 13.9 13.10 Label L_incompatible_class_change_error; 13.11 switch (iid) {
14.1 --- a/src/cpu/x86/vm/runtime_x86_32.cpp Fri Mar 27 10:57:42 2015 +0100 14.2 +++ b/src/cpu/x86/vm/runtime_x86_32.cpp Fri May 29 10:58:45 2015 +0200 14.3 @@ -126,10 +126,6 @@ 14.4 14.5 // rax: exception handler for given <exception oop/exception pc> 14.6 14.7 - // Restore SP from BP if the exception PC is a MethodHandle call site. 14.8 - __ cmpl(Address(rcx, JavaThread::is_method_handle_return_offset()), 0); 14.9 - __ cmovptr(Assembler::notEqual, rsp, rbp_mh_SP_save); 14.10 - 14.11 // We have a handler in rax, (could be deopt blob) 14.12 // rdx - throwing pc, deopt blob will need it. 14.13
15.1 --- a/src/cpu/x86/vm/sharedRuntime_x86_64.cpp Fri Mar 27 10:57:42 2015 +0100 15.2 +++ b/src/cpu/x86/vm/sharedRuntime_x86_64.cpp Fri May 29 10:58:45 2015 +0200 15.3 @@ -4017,8 +4017,8 @@ 15.4 15.5 // Save callee-saved registers. See x86_64.ad. 15.6 15.7 - // rbp is an implicitly saved callee saved register (i.e. the calling 15.8 - // convention will save restore it in prolog/epilog) Other than that 15.9 + // rbp is an implicitly saved callee saved register (i.e., the calling 15.10 + // convention will save/restore it in the prolog/epilog). Other than that 15.11 // there are no callee save registers now that adapter frames are gone. 15.12 15.13 __ movptr(Address(rsp, SimpleRuntimeFrame::rbp_off << LogBytesPerInt), rbp); 15.14 @@ -4060,9 +4060,9 @@ 15.15 15.16 // Restore callee-saved registers 15.17 15.18 - // rbp is an implicitly saved callee saved register (i.e. the calling 15.19 + // rbp is an implicitly saved callee-saved register (i.e., the calling 15.20 // convention will save restore it in prolog/epilog) Other than that 15.21 - // there are no callee save registers no that adapter frames are gone. 15.22 + // there are no callee save registers now that adapter frames are gone. 15.23 15.24 __ movptr(rbp, Address(rsp, SimpleRuntimeFrame::rbp_off << LogBytesPerInt)); 15.25 15.26 @@ -4071,10 +4071,6 @@ 15.27 15.28 // rax: exception handler 15.29 15.30 - // Restore SP from BP if the exception PC is a MethodHandle call site. 15.31 - __ cmpl(Address(r15_thread, JavaThread::is_method_handle_return_offset()), 0); 15.32 - __ cmovptr(Assembler::notEqual, rsp, rbp_mh_SP_save); 15.33 - 15.34 // We have a handler in rax (could be deopt blob). 15.35 __ mov(r8, rax); 15.36
16.1 --- a/src/cpu/x86/vm/x86.ad Fri Mar 27 10:57:42 2015 +0100 16.2 +++ b/src/cpu/x86/vm/x86.ad Fri May 29 10:58:45 2015 +0200 16.3 @@ -912,21 +912,6 @@ 16.4 16.5 encode %{ 16.6 16.7 - enc_class preserve_SP %{ 16.8 - debug_only(int off0 = cbuf.insts_size()); 16.9 - MacroAssembler _masm(&cbuf); 16.10 - // RBP is preserved across all calls, even compiled calls. 16.11 - // Use it to preserve RSP in places where the callee might change the SP. 16.12 - __ movptr(rbp_mh_SP_save, rsp); 16.13 - debug_only(int off1 = cbuf.insts_size()); 16.14 - assert(off1 - off0 == preserve_SP_size(), "correct size prediction"); 16.15 - %} 16.16 - 16.17 - enc_class restore_SP %{ 16.18 - MacroAssembler _masm(&cbuf); 16.19 - __ movptr(rsp, rbp_mh_SP_save); 16.20 - %} 16.21 - 16.22 enc_class call_epilog %{ 16.23 if (VerifyStackAtCalls) { 16.24 // Check that stack depth is unchanged: find majik cookie on stack
17.1 --- a/src/cpu/x86/vm/x86_32.ad Fri Mar 27 10:57:42 2015 +0100 17.2 +++ b/src/cpu/x86/vm/x86_32.ad Fri May 29 10:58:45 2015 +0200 17.3 @@ -123,50 +123,94 @@ 17.4 // 2) reg_class interpreter_method_oop_reg ( /* as def'd in frame section */ ) 17.5 // 3) reg_class stack_slots( /* one chunk of stack-based "registers" */ ) 17.6 // 17.7 +// Class for no registers (empty set). 17.8 +reg_class no_reg(); 17.9 + 17.10 // Class for all registers 17.11 -reg_class any_reg(EAX, EDX, EBP, EDI, ESI, ECX, EBX, ESP); 17.12 +reg_class any_reg_with_ebp(EAX, EDX, EBP, EDI, ESI, ECX, EBX, ESP); 17.13 +// Class for all registers (excluding EBP) 17.14 +reg_class any_reg_no_ebp(EAX, EDX, EDI, ESI, ECX, EBX, ESP); 17.15 +// Dynamic register class that selects at runtime between register classes 17.16 +// any_reg and any_no_ebp_reg (depending on the value of the flag PreserveFramePointer). 17.17 +// Equivalent to: return PreserveFramePointer ? any_no_ebp_reg : any_reg; 17.18 +reg_class_dynamic any_reg(any_reg_no_ebp, any_reg_with_ebp, %{ PreserveFramePointer %}); 17.19 + 17.20 // Class for general registers 17.21 -reg_class int_reg(EAX, EDX, EBP, EDI, ESI, ECX, EBX); 17.22 -// Class for general registers which may be used for implicit null checks on win95 17.23 -// Also safe for use by tailjump. We don't want to allocate in rbp, 17.24 -reg_class int_reg_no_rbp(EAX, EDX, EDI, ESI, ECX, EBX); 17.25 +reg_class int_reg_with_ebp(EAX, EDX, EBP, EDI, ESI, ECX, EBX); 17.26 +// Class for general registers (excluding EBP). 17.27 +// This register class can be used for implicit null checks on win95. 17.28 +// It is also safe for use by tailjumps (we don't want to allocate in ebp). 17.29 +// Used also if the PreserveFramePointer flag is true. 17.30 +reg_class int_reg_no_ebp(EAX, EDX, EDI, ESI, ECX, EBX); 17.31 +// Dynamic register class that selects between int_reg and int_reg_no_ebp. 17.32 +reg_class_dynamic int_reg(int_reg_no_ebp, int_reg_with_ebp, %{ PreserveFramePointer %}); 17.33 + 17.34 // Class of "X" registers 17.35 reg_class int_x_reg(EBX, ECX, EDX, EAX); 17.36 + 17.37 // Class of registers that can appear in an address with no offset. 17.38 // EBP and ESP require an extra instruction byte for zero offset. 17.39 // Used in fast-unlock 17.40 reg_class p_reg(EDX, EDI, ESI, EBX); 17.41 -// Class for general registers not including ECX 17.42 -reg_class ncx_reg(EAX, EDX, EBP, EDI, ESI, EBX); 17.43 -// Class for general registers not including EAX 17.44 + 17.45 +// Class for general registers excluding ECX 17.46 +reg_class ncx_reg_with_ebp(EAX, EDX, EBP, EDI, ESI, EBX); 17.47 +// Class for general registers excluding ECX (and EBP) 17.48 +reg_class ncx_reg_no_ebp(EAX, EDX, EDI, ESI, EBX); 17.49 +// Dynamic register class that selects between ncx_reg and ncx_reg_no_ebp. 17.50 +reg_class_dynamic ncx_reg(ncx_reg_no_ebp, ncx_reg_with_ebp, %{ PreserveFramePointer %}); 17.51 + 17.52 +// Class for general registers excluding EAX 17.53 reg_class nax_reg(EDX, EDI, ESI, ECX, EBX); 17.54 -// Class for general registers not including EAX or EBX. 17.55 -reg_class nabx_reg(EDX, EDI, ESI, ECX, EBP); 17.56 + 17.57 +// Class for general registers excluding EAX and EBX. 17.58 +reg_class nabx_reg_with_ebp(EDX, EDI, ESI, ECX, EBP); 17.59 +// Class for general registers excluding EAX and EBX (and EBP) 17.60 +reg_class nabx_reg_no_ebp(EDX, EDI, ESI, ECX); 17.61 +// Dynamic register class that selects between nabx_reg and nabx_reg_no_ebp. 17.62 +reg_class_dynamic nabx_reg(nabx_reg_no_ebp, nabx_reg_with_ebp, %{ PreserveFramePointer %}); 17.63 + 17.64 // Class of EAX (for multiply and divide operations) 17.65 reg_class eax_reg(EAX); 17.66 + 17.67 // Class of EBX (for atomic add) 17.68 reg_class ebx_reg(EBX); 17.69 + 17.70 // Class of ECX (for shift and JCXZ operations and cmpLTMask) 17.71 reg_class ecx_reg(ECX); 17.72 + 17.73 // Class of EDX (for multiply and divide operations) 17.74 reg_class edx_reg(EDX); 17.75 + 17.76 // Class of EDI (for synchronization) 17.77 reg_class edi_reg(EDI); 17.78 + 17.79 // Class of ESI (for synchronization) 17.80 reg_class esi_reg(ESI); 17.81 -// Singleton class for interpreter's stack pointer 17.82 -reg_class ebp_reg(EBP); 17.83 + 17.84 // Singleton class for stack pointer 17.85 reg_class sp_reg(ESP); 17.86 + 17.87 // Singleton class for instruction pointer 17.88 // reg_class ip_reg(EIP); 17.89 + 17.90 // Class of integer register pairs 17.91 -reg_class long_reg( EAX,EDX, ECX,EBX, EBP,EDI ); 17.92 +reg_class long_reg_with_ebp( EAX,EDX, ECX,EBX, EBP,EDI ); 17.93 +// Class of integer register pairs (excluding EBP and EDI); 17.94 +reg_class long_reg_no_ebp( EAX,EDX, ECX,EBX ); 17.95 +// Dynamic register class that selects between long_reg and long_reg_no_ebp. 17.96 +reg_class_dynamic long_reg(long_reg_no_ebp, long_reg_with_ebp, %{ PreserveFramePointer %}); 17.97 + 17.98 // Class of integer register pairs that aligns with calling convention 17.99 reg_class eadx_reg( EAX,EDX ); 17.100 reg_class ebcx_reg( ECX,EBX ); 17.101 + 17.102 // Not AX or DX, used in divides 17.103 -reg_class nadx_reg( EBX,ECX,ESI,EDI,EBP ); 17.104 +reg_class nadx_reg_with_ebp(EBX, ECX, ESI, EDI, EBP); 17.105 +// Not AX or DX (and neither EBP), used in divides 17.106 +reg_class nadx_reg_no_ebp(EBX, ECX, ESI, EDI); 17.107 +// Dynamic register class that selects between nadx_reg and nadx_reg_no_ebp. 17.108 +reg_class_dynamic nadx_reg(nadx_reg_no_ebp, nadx_reg_with_ebp, %{ PreserveFramePointer %}); 17.109 17.110 // Floating point registers. Notice FPR0 is not a choice. 17.111 // FPR0 is not ever allocated; we use clever encodings to fake 17.112 @@ -240,18 +284,11 @@ 17.113 return size; 17.114 } 17.115 17.116 -static int preserve_SP_size() { 17.117 - return 2; // op, rm(reg/reg) 17.118 -} 17.119 - 17.120 // !!!!! Special hack to get all type of calls to specify the byte offset 17.121 // from the start of the call to the point where the return address 17.122 // will point. 17.123 int MachCallStaticJavaNode::ret_addr_offset() { 17.124 - int offset = 5 + pre_call_resets_size(); // 5 bytes from start of call to where return address points 17.125 - if (_method_handle_invoke) 17.126 - offset += preserve_SP_size(); 17.127 - return offset; 17.128 + return 5 + pre_call_resets_size(); // 5 bytes from start of call to where return address points 17.129 } 17.130 17.131 int MachCallDynamicJavaNode::ret_addr_offset() { 17.132 @@ -285,15 +322,6 @@ 17.133 17.134 // The address of the call instruction needs to be 4-byte aligned to 17.135 // ensure that it does not span a cache line so that it can be patched. 17.136 -int CallStaticJavaHandleNode::compute_padding(int current_offset) const { 17.137 - current_offset += pre_call_resets_size(); // skip fldcw, if any 17.138 - current_offset += preserve_SP_size(); // skip mov rbp, rsp 17.139 - current_offset += 1; // skip call opcode byte 17.140 - return round_to(current_offset, alignment_required()) - current_offset; 17.141 -} 17.142 - 17.143 -// The address of the call instruction needs to be 4-byte aligned to 17.144 -// ensure that it does not span a cache line so that it can be patched. 17.145 int CallDynamicJavaDirectNode::compute_padding(int current_offset) const { 17.146 current_offset += pre_call_resets_size(); // skip fldcw, if any 17.147 current_offset += 5; // skip MOV instruction 17.148 @@ -523,6 +551,10 @@ 17.149 st->print("# stack bang (%d bytes)", bangsize); 17.150 st->print("\n\t"); 17.151 st->print("PUSH EBP\t# Save EBP"); 17.152 + if (PreserveFramePointer) { 17.153 + st->print("\n\t"); 17.154 + st->print("MOV EBP, ESP\t# Save the caller's SP into EBP"); 17.155 + } 17.156 if (framesize) { 17.157 st->print("\n\t"); 17.158 st->print("SUB ESP, #%d\t# Create frame",framesize); 17.159 @@ -532,6 +564,10 @@ 17.160 st->print("\n\t"); 17.161 framesize -= wordSize; 17.162 st->print("MOV [ESP + #%d], EBP\t# Save EBP",framesize); 17.163 + if (PreserveFramePointer) { 17.164 + st->print("\n\t"); 17.165 + st->print("MOV EBP, [ESP + #%d]\t# Save the caller's SP into EBP", (framesize + wordSize)); 17.166 + } 17.167 } 17.168 17.169 if (VerifyStackAtCalls) { 17.170 @@ -1488,7 +1524,7 @@ 17.171 } 17.172 17.173 const RegMask Matcher::method_handle_invoke_SP_save_mask() { 17.174 - return EBP_REG_mask(); 17.175 + return NO_REG_mask(); 17.176 } 17.177 17.178 // Returns true if the high 32 bits of the value is known to be zero. 17.179 @@ -3734,7 +3770,7 @@ 17.180 17.181 // On windows95, EBP is not safe to use for implicit null tests. 17.182 operand eRegP_no_EBP() %{ 17.183 - constraint(ALLOC_IN_RC(int_reg_no_rbp)); 17.184 + constraint(ALLOC_IN_RC(int_reg_no_ebp)); 17.185 match(RegP); 17.186 match(eAXRegP); 17.187 match(eBXRegP); 17.188 @@ -3823,13 +3859,6 @@ 17.189 interface(REG_INTER); 17.190 %} 17.191 17.192 -operand eBPRegP() %{ 17.193 - constraint(ALLOC_IN_RC(ebp_reg)); 17.194 - match(RegP); 17.195 - format %{ "EBP" %} 17.196 - interface(REG_INTER); 17.197 -%} 17.198 - 17.199 operand eRegL() %{ 17.200 constraint(ALLOC_IN_RC(long_reg)); 17.201 match(RegL); 17.202 @@ -12708,7 +12737,6 @@ 17.203 // compute_padding() functions will have to be adjusted. 17.204 instruct CallStaticJavaDirect(method meth) %{ 17.205 match(CallStaticJava); 17.206 - predicate(! ((CallStaticJavaNode*)n)->is_method_handle_invoke()); 17.207 effect(USE meth); 17.208 17.209 ins_cost(300); 17.210 @@ -12722,29 +12750,6 @@ 17.211 ins_alignment(4); 17.212 %} 17.213 17.214 -// Call Java Static Instruction (method handle version) 17.215 -// Note: If this code changes, the corresponding ret_addr_offset() and 17.216 -// compute_padding() functions will have to be adjusted. 17.217 -instruct CallStaticJavaHandle(method meth, eBPRegP ebp_mh_SP_save) %{ 17.218 - match(CallStaticJava); 17.219 - predicate(((CallStaticJavaNode*)n)->is_method_handle_invoke()); 17.220 - effect(USE meth); 17.221 - // EBP is saved by all callees (for interpreter stack correction). 17.222 - // We use it here for a similar purpose, in {preserve,restore}_SP. 17.223 - 17.224 - ins_cost(300); 17.225 - format %{ "CALL,static/MethodHandle " %} 17.226 - opcode(0xE8); /* E8 cd */ 17.227 - ins_encode( pre_call_resets, 17.228 - preserve_SP, 17.229 - Java_Static_Call( meth ), 17.230 - restore_SP, 17.231 - call_epilog, 17.232 - post_call_FPU ); 17.233 - ins_pipe( pipe_slow ); 17.234 - ins_alignment(4); 17.235 -%} 17.236 - 17.237 // Call Java Dynamic Instruction 17.238 // Note: If this code changes, the corresponding ret_addr_offset() and 17.239 // compute_padding() functions will have to be adjusted.
18.1 --- a/src/cpu/x86/vm/x86_64.ad Fri Mar 27 10:57:42 2015 +0100 18.2 +++ b/src/cpu/x86/vm/x86_64.ad Fri May 29 10:58:45 2015 +0200 18.3 @@ -166,42 +166,67 @@ 18.4 // 3) reg_class stack_slots( /* one chunk of stack-based "registers" */ ) 18.5 // 18.6 18.7 -// Class for all pointer registers (including RSP) 18.8 -reg_class any_reg(RAX, RAX_H, 18.9 - RDX, RDX_H, 18.10 - RBP, RBP_H, 18.11 - RDI, RDI_H, 18.12 - RSI, RSI_H, 18.13 - RCX, RCX_H, 18.14 - RBX, RBX_H, 18.15 - RSP, RSP_H, 18.16 - R8, R8_H, 18.17 - R9, R9_H, 18.18 - R10, R10_H, 18.19 - R11, R11_H, 18.20 - R12, R12_H, 18.21 - R13, R13_H, 18.22 - R14, R14_H, 18.23 - R15, R15_H); 18.24 - 18.25 -// Class for all pointer registers except RSP 18.26 -reg_class ptr_reg(RAX, RAX_H, 18.27 - RDX, RDX_H, 18.28 - RBP, RBP_H, 18.29 - RDI, RDI_H, 18.30 - RSI, RSI_H, 18.31 - RCX, RCX_H, 18.32 - RBX, RBX_H, 18.33 - R8, R8_H, 18.34 - R9, R9_H, 18.35 - R10, R10_H, 18.36 - R11, R11_H, 18.37 - R13, R13_H, 18.38 - R14, R14_H); 18.39 - 18.40 -// Class for all pointer registers except RAX and RSP 18.41 -reg_class ptr_no_rax_reg(RDX, RDX_H, 18.42 - RBP, RBP_H, 18.43 +// Empty register class. 18.44 +reg_class no_reg(); 18.45 + 18.46 +// Class for all pointer registers (including RSP and RBP) 18.47 +reg_class any_reg_with_rbp(RAX, RAX_H, 18.48 + RDX, RDX_H, 18.49 + RBP, RBP_H, 18.50 + RDI, RDI_H, 18.51 + RSI, RSI_H, 18.52 + RCX, RCX_H, 18.53 + RBX, RBX_H, 18.54 + RSP, RSP_H, 18.55 + R8, R8_H, 18.56 + R9, R9_H, 18.57 + R10, R10_H, 18.58 + R11, R11_H, 18.59 + R12, R12_H, 18.60 + R13, R13_H, 18.61 + R14, R14_H, 18.62 + R15, R15_H); 18.63 + 18.64 +// Class for all pointer registers (including RSP, but excluding RBP) 18.65 +reg_class any_reg_no_rbp(RAX, RAX_H, 18.66 + RDX, RDX_H, 18.67 + RDI, RDI_H, 18.68 + RSI, RSI_H, 18.69 + RCX, RCX_H, 18.70 + RBX, RBX_H, 18.71 + RSP, RSP_H, 18.72 + R8, R8_H, 18.73 + R9, R9_H, 18.74 + R10, R10_H, 18.75 + R11, R11_H, 18.76 + R12, R12_H, 18.77 + R13, R13_H, 18.78 + R14, R14_H, 18.79 + R15, R15_H); 18.80 + 18.81 +// Dynamic register class that selects at runtime between register classes 18.82 +// any_reg_no_rbp and any_reg_with_rbp (depending on the value of the flag PreserveFramePointer). 18.83 +// Equivalent to: return PreserveFramePointer ? any_reg_no_rbp : any_reg_with_rbp; 18.84 +reg_class_dynamic any_reg(any_reg_no_rbp, any_reg_with_rbp, %{ PreserveFramePointer %}); 18.85 + 18.86 +// Class for all pointer registers (excluding RSP) 18.87 +reg_class ptr_reg_with_rbp(RAX, RAX_H, 18.88 + RDX, RDX_H, 18.89 + RBP, RBP_H, 18.90 + RDI, RDI_H, 18.91 + RSI, RSI_H, 18.92 + RCX, RCX_H, 18.93 + RBX, RBX_H, 18.94 + R8, R8_H, 18.95 + R9, R9_H, 18.96 + R10, R10_H, 18.97 + R11, R11_H, 18.98 + R13, R13_H, 18.99 + R14, R14_H); 18.100 + 18.101 +// Class for all pointer registers (excluding RSP and RBP) 18.102 +reg_class ptr_reg_no_rbp(RAX, RAX_H, 18.103 + RDX, RDX_H, 18.104 RDI, RDI_H, 18.105 RSI, RSI_H, 18.106 RCX, RCX_H, 18.107 @@ -213,31 +238,66 @@ 18.108 R13, R13_H, 18.109 R14, R14_H); 18.110 18.111 -reg_class ptr_no_rbp_reg(RDX, RDX_H, 18.112 - RAX, RAX_H, 18.113 - RDI, RDI_H, 18.114 - RSI, RSI_H, 18.115 - RCX, RCX_H, 18.116 - RBX, RBX_H, 18.117 - R8, R8_H, 18.118 - R9, R9_H, 18.119 - R10, R10_H, 18.120 - R11, R11_H, 18.121 - R13, R13_H, 18.122 - R14, R14_H); 18.123 - 18.124 -// Class for all pointer registers except RAX, RBX and RSP 18.125 -reg_class ptr_no_rax_rbx_reg(RDX, RDX_H, 18.126 - RBP, RBP_H, 18.127 - RDI, RDI_H, 18.128 - RSI, RSI_H, 18.129 - RCX, RCX_H, 18.130 - R8, R8_H, 18.131 - R9, R9_H, 18.132 - R10, R10_H, 18.133 - R11, R11_H, 18.134 - R13, R13_H, 18.135 - R14, R14_H); 18.136 +// Dynamic register class that selects between ptr_reg_no_rbp and ptr_reg_with_rbp. 18.137 +reg_class_dynamic ptr_reg(ptr_reg_no_rbp, ptr_reg_with_rbp, %{ PreserveFramePointer %}); 18.138 + 18.139 +// Class for all pointer registers (excluding RAX and RSP) 18.140 +reg_class ptr_no_rax_reg_with_rbp(RDX, RDX_H, 18.141 + RBP, RBP_H, 18.142 + RDI, RDI_H, 18.143 + RSI, RSI_H, 18.144 + RCX, RCX_H, 18.145 + RBX, RBX_H, 18.146 + R8, R8_H, 18.147 + R9, R9_H, 18.148 + R10, R10_H, 18.149 + R11, R11_H, 18.150 + R13, R13_H, 18.151 + R14, R14_H); 18.152 + 18.153 +// Class for all pointer registers (excluding RAX, RSP, and RBP) 18.154 +reg_class ptr_no_rax_reg_no_rbp(RDX, RDX_H, 18.155 + RDI, RDI_H, 18.156 + RSI, RSI_H, 18.157 + RCX, RCX_H, 18.158 + RBX, RBX_H, 18.159 + R8, R8_H, 18.160 + R9, R9_H, 18.161 + R10, R10_H, 18.162 + R11, R11_H, 18.163 + R13, R13_H, 18.164 + R14, R14_H); 18.165 + 18.166 +// Dynamic register class that selects between ptr_no_rax_reg_no_rbp and ptr_no_rax_reg_with_rbp. 18.167 +reg_class_dynamic ptr_no_rax_reg(ptr_no_rax_reg_no_rbp, ptr_no_rax_reg_with_rbp, %{ PreserveFramePointer %}); 18.168 + 18.169 +// Class for all pointer registers (excluding RAX, RBX, and RSP) 18.170 +reg_class ptr_no_rax_rbx_reg_with_rbp(RDX, RDX_H, 18.171 + RBP, RBP_H, 18.172 + RDI, RDI_H, 18.173 + RSI, RSI_H, 18.174 + RCX, RCX_H, 18.175 + R8, R8_H, 18.176 + R9, R9_H, 18.177 + R10, R10_H, 18.178 + R11, R11_H, 18.179 + R13, R13_H, 18.180 + R14, R14_H); 18.181 + 18.182 +// Class for all pointer registers (excluding RAX, RBX, RSP, and RBP) 18.183 +reg_class ptr_no_rax_rbx_reg_no_rbp(RDX, RDX_H, 18.184 + RDI, RDI_H, 18.185 + RSI, RSI_H, 18.186 + RCX, RCX_H, 18.187 + R8, R8_H, 18.188 + R9, R9_H, 18.189 + R10, R10_H, 18.190 + R11, R11_H, 18.191 + R13, R13_H, 18.192 + R14, R14_H); 18.193 + 18.194 +// Dynamic register class that selects between ptr_no_rax_rbx_reg_no_rbp and ptr_no_rax_rbx_reg_with_rbp. 18.195 +reg_class_dynamic ptr_no_rax_rbx_reg(ptr_no_rax_rbx_reg_no_rbp, ptr_no_rax_rbx_reg_with_rbp, %{ PreserveFramePointer %}); 18.196 18.197 // Singleton class for RAX pointer register 18.198 reg_class ptr_rax_reg(RAX, RAX_H); 18.199 @@ -251,59 +311,29 @@ 18.200 // Singleton class for RDI pointer register 18.201 reg_class ptr_rdi_reg(RDI, RDI_H); 18.202 18.203 -// Singleton class for RBP pointer register 18.204 -reg_class ptr_rbp_reg(RBP, RBP_H); 18.205 - 18.206 // Singleton class for stack pointer 18.207 reg_class ptr_rsp_reg(RSP, RSP_H); 18.208 18.209 // Singleton class for TLS pointer 18.210 reg_class ptr_r15_reg(R15, R15_H); 18.211 18.212 -// Class for all long registers (except RSP) 18.213 -reg_class long_reg(RAX, RAX_H, 18.214 - RDX, RDX_H, 18.215 - RBP, RBP_H, 18.216 - RDI, RDI_H, 18.217 - RSI, RSI_H, 18.218 - RCX, RCX_H, 18.219 - RBX, RBX_H, 18.220 - R8, R8_H, 18.221 - R9, R9_H, 18.222 - R10, R10_H, 18.223 - R11, R11_H, 18.224 - R13, R13_H, 18.225 - R14, R14_H); 18.226 - 18.227 -// Class for all long registers except RAX, RDX (and RSP) 18.228 -reg_class long_no_rax_rdx_reg(RBP, RBP_H, 18.229 - RDI, RDI_H, 18.230 - RSI, RSI_H, 18.231 - RCX, RCX_H, 18.232 - RBX, RBX_H, 18.233 - R8, R8_H, 18.234 - R9, R9_H, 18.235 - R10, R10_H, 18.236 - R11, R11_H, 18.237 - R13, R13_H, 18.238 - R14, R14_H); 18.239 - 18.240 -// Class for all long registers except RCX (and RSP) 18.241 -reg_class long_no_rcx_reg(RBP, RBP_H, 18.242 - RDI, RDI_H, 18.243 - RSI, RSI_H, 18.244 - RAX, RAX_H, 18.245 - RDX, RDX_H, 18.246 - RBX, RBX_H, 18.247 - R8, R8_H, 18.248 - R9, R9_H, 18.249 - R10, R10_H, 18.250 - R11, R11_H, 18.251 - R13, R13_H, 18.252 - R14, R14_H); 18.253 - 18.254 -// Class for all long registers except RAX (and RSP) 18.255 -reg_class long_no_rax_reg(RBP, RBP_H, 18.256 +// Class for all long registers (excluding RSP) 18.257 +reg_class long_reg_with_rbp(RAX, RAX_H, 18.258 + RDX, RDX_H, 18.259 + RBP, RBP_H, 18.260 + RDI, RDI_H, 18.261 + RSI, RSI_H, 18.262 + RCX, RCX_H, 18.263 + RBX, RBX_H, 18.264 + R8, R8_H, 18.265 + R9, R9_H, 18.266 + R10, R10_H, 18.267 + R11, R11_H, 18.268 + R13, R13_H, 18.269 + R14, R14_H); 18.270 + 18.271 +// Class for all long registers (excluding RSP and RBP) 18.272 +reg_class long_reg_no_rbp(RAX, RAX_H, 18.273 RDX, RDX_H, 18.274 RDI, RDI_H, 18.275 RSI, RSI_H, 18.276 @@ -316,6 +346,67 @@ 18.277 R13, R13_H, 18.278 R14, R14_H); 18.279 18.280 +// Dynamic register class that selects between long_reg_no_rbp and long_reg_with_rbp. 18.281 +reg_class_dynamic long_reg(long_reg_no_rbp, long_reg_with_rbp, %{ PreserveFramePointer %}); 18.282 + 18.283 +// Class for all long registers (excluding RAX, RDX and RSP) 18.284 +reg_class long_no_rax_rdx_reg_with_rbp(RBP, RBP_H, 18.285 + RDI, RDI_H, 18.286 + RSI, RSI_H, 18.287 + RCX, RCX_H, 18.288 + RBX, RBX_H, 18.289 + R8, R8_H, 18.290 + R9, R9_H, 18.291 + R10, R10_H, 18.292 + R11, R11_H, 18.293 + R13, R13_H, 18.294 + R14, R14_H); 18.295 + 18.296 +// Class for all long registers (excluding RAX, RDX, RSP, and RBP) 18.297 +reg_class long_no_rax_rdx_reg_no_rbp(RDI, RDI_H, 18.298 + RSI, RSI_H, 18.299 + RCX, RCX_H, 18.300 + RBX, RBX_H, 18.301 + R8, R8_H, 18.302 + R9, R9_H, 18.303 + R10, R10_H, 18.304 + R11, R11_H, 18.305 + R13, R13_H, 18.306 + R14, R14_H); 18.307 + 18.308 +// Dynamic register class that selects between long_no_rax_rdx_reg_no_rbp and long_no_rax_rdx_reg_with_rbp. 18.309 +reg_class_dynamic long_no_rax_rdx_reg(long_no_rax_rdx_reg_no_rbp, long_no_rax_rdx_reg_with_rbp, %{ PreserveFramePointer %}); 18.310 + 18.311 +// Class for all long registers (excluding RCX and RSP) 18.312 +reg_class long_no_rcx_reg_with_rbp(RBP, RBP_H, 18.313 + RDI, RDI_H, 18.314 + RSI, RSI_H, 18.315 + RAX, RAX_H, 18.316 + RDX, RDX_H, 18.317 + RBX, RBX_H, 18.318 + R8, R8_H, 18.319 + R9, R9_H, 18.320 + R10, R10_H, 18.321 + R11, R11_H, 18.322 + R13, R13_H, 18.323 + R14, R14_H); 18.324 + 18.325 +// Class for all long registers (excluding RCX, RSP, and RBP) 18.326 +reg_class long_no_rcx_reg_no_rbp(RDI, RDI_H, 18.327 + RSI, RSI_H, 18.328 + RAX, RAX_H, 18.329 + RDX, RDX_H, 18.330 + RBX, RBX_H, 18.331 + R8, R8_H, 18.332 + R9, R9_H, 18.333 + R10, R10_H, 18.334 + R11, R11_H, 18.335 + R13, R13_H, 18.336 + R14, R14_H); 18.337 + 18.338 +// Dynamic register class that selects between long_no_rcx_reg_no_rbp and long_no_rcx_reg_with_rbp. 18.339 +reg_class_dynamic long_no_rcx_reg(long_no_rcx_reg_no_rbp, long_no_rcx_reg_with_rbp, %{ PreserveFramePointer %}); 18.340 + 18.341 // Singleton class for RAX long register 18.342 reg_class long_rax_reg(RAX, RAX_H); 18.343 18.344 @@ -325,27 +416,27 @@ 18.345 // Singleton class for RDX long register 18.346 reg_class long_rdx_reg(RDX, RDX_H); 18.347 18.348 -// Class for all int registers (except RSP) 18.349 -reg_class int_reg(RAX, 18.350 - RDX, 18.351 - RBP, 18.352 - RDI, 18.353 - RSI, 18.354 - RCX, 18.355 - RBX, 18.356 - R8, 18.357 - R9, 18.358 - R10, 18.359 - R11, 18.360 - R13, 18.361 - R14); 18.362 - 18.363 -// Class for all int registers except RCX (and RSP) 18.364 -reg_class int_no_rcx_reg(RAX, 18.365 +// Class for all int registers (excluding RSP) 18.366 +reg_class int_reg_with_rbp(RAX, 18.367 + RDX, 18.368 + RBP, 18.369 + RDI, 18.370 + RSI, 18.371 + RCX, 18.372 + RBX, 18.373 + R8, 18.374 + R9, 18.375 + R10, 18.376 + R11, 18.377 + R13, 18.378 + R14); 18.379 + 18.380 +// Class for all int registers (excluding RSP and RBP) 18.381 +reg_class int_reg_no_rbp(RAX, 18.382 RDX, 18.383 - RBP, 18.384 RDI, 18.385 RSI, 18.386 + RCX, 18.387 RBX, 18.388 R8, 18.389 R9, 18.390 @@ -354,18 +445,66 @@ 18.391 R13, 18.392 R14); 18.393 18.394 -// Class for all int registers except RAX, RDX (and RSP) 18.395 -reg_class int_no_rax_rdx_reg(RBP, 18.396 - RDI, 18.397 - RSI, 18.398 - RCX, 18.399 - RBX, 18.400 - R8, 18.401 - R9, 18.402 - R10, 18.403 - R11, 18.404 - R13, 18.405 - R14); 18.406 +// Dynamic register class that selects between int_reg_no_rbp and int_reg_with_rbp. 18.407 +reg_class_dynamic int_reg(int_reg_no_rbp, int_reg_with_rbp, %{ PreserveFramePointer %}); 18.408 + 18.409 +// Class for all int registers (excluding RCX and RSP) 18.410 +reg_class int_no_rcx_reg_with_rbp(RAX, 18.411 + RDX, 18.412 + RBP, 18.413 + RDI, 18.414 + RSI, 18.415 + RBX, 18.416 + R8, 18.417 + R9, 18.418 + R10, 18.419 + R11, 18.420 + R13, 18.421 + R14); 18.422 + 18.423 +// Class for all int registers (excluding RCX, RSP, and RBP) 18.424 +reg_class int_no_rcx_reg_no_rbp(RAX, 18.425 + RDX, 18.426 + RDI, 18.427 + RSI, 18.428 + RBX, 18.429 + R8, 18.430 + R9, 18.431 + R10, 18.432 + R11, 18.433 + R13, 18.434 + R14); 18.435 + 18.436 +// Dynamic register class that selects between int_no_rcx_reg_no_rbp and int_no_rcx_reg_with_rbp. 18.437 +reg_class_dynamic int_no_rcx_reg(int_no_rcx_reg_no_rbp, int_no_rcx_reg_with_rbp, %{ PreserveFramePointer %}); 18.438 + 18.439 +// Class for all int registers (excluding RAX, RDX, and RSP) 18.440 +reg_class int_no_rax_rdx_reg_with_rbp(RBP, 18.441 + RDI, 18.442 + RSI, 18.443 + RCX, 18.444 + RBX, 18.445 + R8, 18.446 + R9, 18.447 + R10, 18.448 + R11, 18.449 + R13, 18.450 + R14); 18.451 + 18.452 +// Class for all int registers (excluding RAX, RDX, RSP, and RBP) 18.453 +reg_class int_no_rax_rdx_reg_no_rbp(RDI, 18.454 + RSI, 18.455 + RCX, 18.456 + RBX, 18.457 + R8, 18.458 + R9, 18.459 + R10, 18.460 + R11, 18.461 + R13, 18.462 + R14); 18.463 + 18.464 +// Dynamic register class that selects between int_no_rax_rdx_reg_no_rbp and int_no_rax_rdx_reg_with_rbp. 18.465 +reg_class_dynamic int_no_rax_rdx_reg(int_no_rax_rdx_reg_no_rbp, int_no_rax_rdx_reg_with_rbp, %{ PreserveFramePointer %}); 18.466 18.467 // Singleton class for RAX int register 18.468 reg_class int_rax_reg(RAX); 18.469 @@ -396,9 +535,6 @@ 18.470 18.471 #define __ _masm. 18.472 18.473 -static int preserve_SP_size() { 18.474 - return 3; // rex.w, op, rm(reg/reg) 18.475 -} 18.476 static int clear_avx_size() { 18.477 return (Compile::current()->max_vector_size() > 16) ? 3 : 0; // vzeroupper 18.478 } 18.479 @@ -409,9 +545,7 @@ 18.480 int MachCallStaticJavaNode::ret_addr_offset() 18.481 { 18.482 int offset = 5; // 5 bytes from start of call to where return address points 18.483 - offset += clear_avx_size(); 18.484 - if (_method_handle_invoke) 18.485 - offset += preserve_SP_size(); 18.486 + offset += clear_avx_size(); 18.487 return offset; 18.488 } 18.489 18.490 @@ -450,16 +584,6 @@ 18.491 18.492 // The address of the call instruction needs to be 4-byte aligned to 18.493 // ensure that it does not span a cache line so that it can be patched. 18.494 -int CallStaticJavaHandleNode::compute_padding(int current_offset) const 18.495 -{ 18.496 - current_offset += preserve_SP_size(); // skip mov rbp, rsp 18.497 - current_offset += clear_avx_size(); // skip vzeroupper 18.498 - current_offset += 1; // skip call opcode byte 18.499 - return round_to(current_offset, alignment_required()) - current_offset; 18.500 -} 18.501 - 18.502 -// The address of the call instruction needs to be 4-byte aligned to 18.503 -// ensure that it does not span a cache line so that it can be patched. 18.504 int CallDynamicJavaDirectNode::compute_padding(int current_offset) const 18.505 { 18.506 current_offset += clear_avx_size(); // skip vzeroupper 18.507 @@ -724,6 +848,10 @@ 18.508 st->print("# stack bang (%d bytes)", bangsize); 18.509 st->print("\n\t"); 18.510 st->print("pushq rbp\t# Save rbp"); 18.511 + if (PreserveFramePointer) { 18.512 + st->print("\n\t"); 18.513 + st->print("movq rbp, rsp\t# Save the caller's SP into rbp"); 18.514 + } 18.515 if (framesize) { 18.516 st->print("\n\t"); 18.517 st->print("subq rsp, #%d\t# Create frame",framesize); 18.518 @@ -732,7 +860,11 @@ 18.519 st->print("subq rsp, #%d\t# Create frame",framesize); 18.520 st->print("\n\t"); 18.521 framesize -= wordSize; 18.522 - st->print("movq [rsp + #%d], rbp\t# Save rbp",framesize); 18.523 + st->print("movq [rsp + #%d], rbp\t# Save rbp",framesize); 18.524 + if (PreserveFramePointer) { 18.525 + st->print("\n\t"); 18.526 + st->print("movq rbp, [rsp + #%d]\t# Save the caller's SP into rbp", (framesize + wordSize)); 18.527 + } 18.528 } 18.529 18.530 if (VerifyStackAtCalls) { 18.531 @@ -1598,8 +1730,9 @@ 18.532 return LONG_RDX_REG_mask(); 18.533 } 18.534 18.535 +// Register for saving SP into on method handle invokes. Not used on x86_64. 18.536 const RegMask Matcher::method_handle_invoke_SP_save_mask() { 18.537 - return PTR_RBP_REG_mask(); 18.538 + return NO_REG_mask(); 18.539 } 18.540 18.541 %} 18.542 @@ -3202,7 +3335,7 @@ 18.543 // Pointer Register 18.544 operand any_RegP() 18.545 %{ 18.546 - constraint(ALLOC_IN_RC(any_reg)); 18.547 + constraint(ALLOC_IN_RC(any_reg)); 18.548 match(RegP); 18.549 match(rax_RegP); 18.550 match(rbx_RegP); 18.551 @@ -3224,8 +3357,8 @@ 18.552 match(rbx_RegP); 18.553 match(rdi_RegP); 18.554 match(rsi_RegP); 18.555 - match(rbp_RegP); 18.556 - match(r15_RegP); // See Q&A below about r15_RegP. 18.557 + match(rbp_RegP); // See Q&A below about 18.558 + match(r15_RegP); // r15_RegP and rbp_RegP. 18.559 18.560 format %{ %} 18.561 interface(REG_INTER); 18.562 @@ -3241,11 +3374,14 @@ 18.563 18.564 // Question: Why is r15_RegP (the read-only TLS register) a match for rRegP? 18.565 // Answer: Operand match rules govern the DFA as it processes instruction inputs. 18.566 -// It's fine for an instruction input which expects rRegP to match a r15_RegP. 18.567 +// It's fine for an instruction input that expects rRegP to match a r15_RegP. 18.568 // The output of an instruction is controlled by the allocator, which respects 18.569 // register class masks, not match rules. Unless an instruction mentions 18.570 // r15_RegP or any_RegP explicitly as its output, r15 will not be considered 18.571 // by the allocator as an input. 18.572 +// The same logic applies to rbp_RegP being a match for rRegP: If PreserveFramePointer==true, 18.573 +// the RBP is used as a proper frame pointer and is not included in ptr_reg. As a 18.574 +// result, RBP is not included in the output of the instruction either. 18.575 18.576 operand no_rax_RegP() 18.577 %{ 18.578 @@ -3259,9 +3395,11 @@ 18.579 interface(REG_INTER); 18.580 %} 18.581 18.582 +// This operand is not allowed to use RBP even if 18.583 +// RBP is not used to hold the frame pointer. 18.584 operand no_rbp_RegP() 18.585 %{ 18.586 - constraint(ALLOC_IN_RC(ptr_no_rbp_reg)); 18.587 + constraint(ALLOC_IN_RC(ptr_reg_no_rbp)); 18.588 match(RegP); 18.589 match(rbx_RegP); 18.590 match(rsi_RegP); 18.591 @@ -3338,16 +3476,6 @@ 18.592 interface(REG_INTER); 18.593 %} 18.594 18.595 -operand rbp_RegP() 18.596 -%{ 18.597 - constraint(ALLOC_IN_RC(ptr_rbp_reg)); 18.598 - match(RegP); 18.599 - match(rRegP); 18.600 - 18.601 - format %{ %} 18.602 - interface(REG_INTER); 18.603 -%} 18.604 - 18.605 operand r15_RegP() 18.606 %{ 18.607 constraint(ALLOC_IN_RC(ptr_r15_reg)); 18.608 @@ -11414,7 +11542,6 @@ 18.609 // compute_padding() functions will have to be adjusted. 18.610 instruct CallStaticJavaDirect(method meth) %{ 18.611 match(CallStaticJava); 18.612 - predicate(!((CallStaticJavaNode*) n)->is_method_handle_invoke()); 18.613 effect(USE meth); 18.614 18.615 ins_cost(300); 18.616 @@ -11425,27 +11552,6 @@ 18.617 ins_alignment(4); 18.618 %} 18.619 18.620 -// Call Java Static Instruction (method handle version) 18.621 -// Note: If this code changes, the corresponding ret_addr_offset() and 18.622 -// compute_padding() functions will have to be adjusted. 18.623 -instruct CallStaticJavaHandle(method meth, rbp_RegP rbp_mh_SP_save) %{ 18.624 - match(CallStaticJava); 18.625 - predicate(((CallStaticJavaNode*) n)->is_method_handle_invoke()); 18.626 - effect(USE meth); 18.627 - // RBP is saved by all callees (for interpreter stack correction). 18.628 - // We use it here for a similar purpose, in {preserve,restore}_SP. 18.629 - 18.630 - ins_cost(300); 18.631 - format %{ "call,static/MethodHandle " %} 18.632 - opcode(0xE8); /* E8 cd */ 18.633 - ins_encode(clear_avx, preserve_SP, 18.634 - Java_Static_Call(meth), 18.635 - restore_SP, 18.636 - call_epilog); 18.637 - ins_pipe(pipe_slow); 18.638 - ins_alignment(4); 18.639 -%} 18.640 - 18.641 // Call Java Dynamic Instruction 18.642 // Note: If this code changes, the corresponding ret_addr_offset() and 18.643 // compute_padding() functions will have to be adjusted.
19.1 --- a/src/share/vm/c1/c1_GraphBuilder.cpp Fri Mar 27 10:57:42 2015 +0100 19.2 +++ b/src/share/vm/c1/c1_GraphBuilder.cpp Fri May 29 10:58:45 2015 +0200 19.3 @@ -4064,7 +4064,7 @@ 19.4 ValueType* type = apop()->type(); 19.5 if (type->is_constant()) { 19.6 ciMethod* target = type->as_ObjectType()->constant_value()->as_member_name()->get_vmtarget(); 19.7 - // If the target is another method handle invoke try recursivly to get 19.8 + // If the target is another method handle invoke, try to recursively get 19.9 // a better target. 19.10 if (target->is_method_handle_intrinsic()) { 19.11 if (try_method_handle_inline(target)) {
20.1 --- a/src/share/vm/c1/c1_LIR.cpp Fri Mar 27 10:57:42 2015 +0100 20.2 +++ b/src/share/vm/c1/c1_LIR.cpp Fri May 29 10:58:45 2015 +0200 20.3 @@ -454,7 +454,7 @@ 20.4 //-------------------visits-------------------------- 20.5 20.6 // complete rework of LIR instruction visitor. 20.7 -// The virtual calls for each instruction type is replaced by a big 20.8 +// The virtual call for each instruction type is replaced by a big 20.9 // switch that adds the operands for each instruction 20.10 20.11 void LIR_OpVisitState::visit(LIR_Op* op) { 20.12 @@ -823,7 +823,8 @@ 20.13 } 20.14 20.15 if (opJavaCall->_info) do_info(opJavaCall->_info); 20.16 - if (opJavaCall->is_method_handle_invoke()) { 20.17 + if (FrameMap::method_handle_invoke_SP_save_opr() != LIR_OprFact::illegalOpr && 20.18 + opJavaCall->is_method_handle_invoke()) { 20.19 opJavaCall->_method_handle_invoke_SP_save_opr = FrameMap::method_handle_invoke_SP_save_opr(); 20.20 do_temp(opJavaCall->_method_handle_invoke_SP_save_opr); 20.21 }
21.1 --- a/src/share/vm/c1/c1_LIR.hpp Fri Mar 27 10:57:42 2015 +0100 21.2 +++ b/src/share/vm/c1/c1_LIR.hpp Fri May 29 10:58:45 2015 +0200 21.3 @@ -1216,10 +1216,8 @@ 21.4 // JSR 292 support. 21.5 bool is_invokedynamic() const { return code() == lir_dynamic_call; } 21.6 bool is_method_handle_invoke() const { 21.7 - return 21.8 - method()->is_compiled_lambda_form() // Java-generated adapter 21.9 - || 21.10 - method()->is_method_handle_intrinsic(); // JVM-generated MH intrinsic 21.11 + return method()->is_compiled_lambda_form() || // Java-generated lambda form 21.12 + method()->is_method_handle_intrinsic(); // JVM-generated MH intrinsic 21.13 } 21.14 21.15 intptr_t vtable_offset() const {
22.1 --- a/src/share/vm/c1/c1_LIRGenerator.cpp Fri Mar 27 10:57:42 2015 +0100 22.2 +++ b/src/share/vm/c1/c1_LIRGenerator.cpp Fri May 29 10:58:45 2015 +0200 22.3 @@ -2887,7 +2887,7 @@ 22.4 // g) lock result registers and emit call operation 22.5 // 22.6 // Before issuing a call, we must spill-save all values on stack 22.7 -// that are in caller-save register. "spill-save" moves thos registers 22.8 +// that are in caller-save register. "spill-save" moves those registers 22.9 // either in a free callee-save register or spills them if no free 22.10 // callee save register is available. 22.11 // 22.12 @@ -2895,7 +2895,7 @@ 22.13 // - if invoked between e) and f), we may lock callee save 22.14 // register in "spill-save" that destroys the receiver register 22.15 // before f) is executed 22.16 -// - if we rearange the f) to be earlier, by loading %o0, it 22.17 +// - if we rearrange f) to be earlier (by loading %o0) it 22.18 // may destroy a value on the stack that is currently in %o0 22.19 // and is waiting to be spilled 22.20 // - if we keep the receiver locked while doing spill-save, 22.21 @@ -2928,14 +2928,16 @@ 22.22 assert(receiver->is_illegal() || receiver->is_equal(LIR_Assembler::receiverOpr()), "must match"); 22.23 22.24 // JSR 292 22.25 - // Preserve the SP over MethodHandle call sites. 22.26 + // Preserve the SP over MethodHandle call sites, if needed. 22.27 ciMethod* target = x->target(); 22.28 bool is_method_handle_invoke = (// %%% FIXME: Are both of these relevant? 22.29 target->is_method_handle_intrinsic() || 22.30 target->is_compiled_lambda_form()); 22.31 if (is_method_handle_invoke) { 22.32 info->set_is_method_handle_invoke(true); 22.33 - __ move(FrameMap::stack_pointer(), FrameMap::method_handle_invoke_SP_save_opr()); 22.34 + if(FrameMap::method_handle_invoke_SP_save_opr() != LIR_OprFact::illegalOpr) { 22.35 + __ move(FrameMap::stack_pointer(), FrameMap::method_handle_invoke_SP_save_opr()); 22.36 + } 22.37 } 22.38 22.39 switch (x->code()) { 22.40 @@ -2975,8 +2977,9 @@ 22.41 } 22.42 22.43 // JSR 292 22.44 - // Restore the SP after MethodHandle call sites. 22.45 - if (is_method_handle_invoke) { 22.46 + // Restore the SP after MethodHandle call sites, if needed. 22.47 + if (is_method_handle_invoke 22.48 + && FrameMap::method_handle_invoke_SP_save_opr() != LIR_OprFact::illegalOpr) { 22.49 __ move(FrameMap::method_handle_invoke_SP_save_opr(), FrameMap::stack_pointer()); 22.50 } 22.51
23.1 --- a/src/share/vm/opto/bytecodeInfo.cpp Fri Mar 27 10:57:42 2015 +0100 23.2 +++ b/src/share/vm/opto/bytecodeInfo.cpp Fri May 29 10:58:45 2015 +0200 23.3 @@ -608,11 +608,11 @@ 23.4 } 23.5 int max_inline_level_adjust = 0; 23.6 if (caller_jvms->method() != NULL) { 23.7 - if (caller_jvms->method()->is_compiled_lambda_form()) 23.8 + if (caller_jvms->method()->is_compiled_lambda_form()) { 23.9 max_inline_level_adjust += 1; // don't count actions in MH or indy adapter frames 23.10 - else if (callee_method->is_method_handle_intrinsic() || 23.11 - callee_method->is_compiled_lambda_form()) { 23.12 - max_inline_level_adjust += 1; // don't count method handle calls from java.lang.invoke implem 23.13 + } else if (callee_method->is_method_handle_intrinsic() || 23.14 + callee_method->is_compiled_lambda_form()) { 23.15 + max_inline_level_adjust += 1; // don't count method handle calls from java.lang.invoke implementation 23.16 } 23.17 if (max_inline_level_adjust != 0 && C->print_inlining() && (Verbose || WizardMode)) { 23.18 CompileTask::print_inline_indent(inline_level());
24.1 --- a/src/share/vm/prims/forte.cpp Fri Mar 27 10:57:42 2015 +0100 24.2 +++ b/src/share/vm/prims/forte.cpp Fri May 29 10:58:45 2015 +0200 24.3 @@ -172,8 +172,27 @@ 24.4 // Now do we have a useful PcDesc? 24.5 if (pc_desc == NULL || 24.6 pc_desc->scope_decode_offset() == DebugInformationRecorder::serialized_null) { 24.7 - // No debug information available for this pc 24.8 - // vframeStream would explode if we try and walk the frames. 24.9 + // No debug information is available for this PC. 24.10 + // 24.11 + // vframeStreamCommon::fill_from_frame() will decode the frame depending 24.12 + // on the state of the thread. 24.13 + // 24.14 + // Case #1: If the thread is in Java (state == _thread_in_Java), then 24.15 + // the vframeStreamCommon object will be filled as if the frame were a native 24.16 + // compiled frame. Therefore, no debug information is needed. 24.17 + // 24.18 + // Case #2: If the thread is in any other state, then two steps will be performed: 24.19 + // - if asserts are enabled, found_bad_method_frame() will be called and 24.20 + // the assert in found_bad_method_frame() will be triggered; 24.21 + // - if asserts are disabled, the vframeStreamCommon object will be filled 24.22 + // as if it were a native compiled frame. 24.23 + // 24.24 + // Case (2) is similar to the way interpreter frames are processed in 24.25 + // vframeStreamCommon::fill_from_interpreter_frame in case no valid BCI 24.26 + // was found for an interpreted frame. If asserts are enabled, the assert 24.27 + // in found_bad_method_frame() will be triggered. If asserts are disabled, 24.28 + // the vframeStreamCommon object will be filled afterwards as if the 24.29 + // interpreter were at the point of entering into the method. 24.30 return false; 24.31 } 24.32 24.33 @@ -230,9 +249,10 @@ 24.34 // a valid method. Then again we may have caught an interpreter 24.35 // frame in the middle of construction and the bci field is 24.36 // not yet valid. 24.37 - 24.38 - *method_p = method; 24.39 if (!method->is_valid_method()) return false; 24.40 + *method_p = method; // If the Method* found is invalid, it is 24.41 + // ignored by forte_fill_call_trace_given_top(). 24.42 + // So set method_p only if the Method is valid. 24.43 24.44 intptr_t bcx = fr->interpreter_frame_bcx(); 24.45 24.46 @@ -247,18 +267,33 @@ 24.47 } 24.48 24.49 24.50 -// Determine if 'fr' can be used to find an initial Java frame. 24.51 -// Return false if it can not find a fully decipherable Java frame 24.52 -// (in other words a frame that isn't safe to use in a vframe stream). 24.53 -// Obviously if it can't even find a Java frame false will also be returned. 24.54 +// Determine if a Java frame can be found starting with the frame 'fr'. 24.55 // 24.56 -// If we find a Java frame decipherable or not then by definition we have 24.57 -// identified a method and that will be returned to the caller via method_p. 24.58 -// If we can determine a bci that is returned also. (Hmm is it possible 24.59 -// to return a method and bci and still return false? ) 24.60 +// Check the return value of find_initial_Java_frame and the value of 24.61 +// 'method_p' to decide on how use the results returned by this method. 24.62 // 24.63 -// The initial Java frame we find (if any) is return via initial_frame_p. 24.64 +// If 'method_p' is not NULL, an initial Java frame has been found and 24.65 +// the stack can be walked starting from that initial frame. In this case, 24.66 +// 'method_p' points to the Method that the initial frame belongs to and 24.67 +// the initial Java frame is returned in initial_frame_p. 24.68 // 24.69 +// find_initial_Java_frame() returns true if a Method has been found (i.e., 24.70 +// 'method_p' is not NULL) and the initial frame that belongs to that Method 24.71 +// is decipherable. 24.72 +// 24.73 +// A frame is considered to be decipherable: 24.74 +// 24.75 +// - if the frame is a compiled frame and a PCDesc is available; 24.76 +// 24.77 +// - if the frame is an interpreter frame that is valid or the thread is 24.78 +// state (_thread_in_native || state == _thread_in_vm || state == _thread_blocked). 24.79 +// 24.80 +// Note that find_initial_Java_frame() can return false even if an initial 24.81 +// Java method was found (e.g., there is no PCDesc available for the method). 24.82 +// 24.83 +// If 'method_p' is NULL, it was not possible to find a Java frame when 24.84 +// walking the stack starting from 'fr'. In this case find_initial_Java_frame 24.85 +// returns false. 24.86 24.87 static bool find_initial_Java_frame(JavaThread* thread, 24.88 frame* fr, 24.89 @@ -278,8 +313,6 @@ 24.90 // recognizable to us. This should only happen if we are in a JRT_LEAF 24.91 // or something called by a JRT_LEAF method. 24.92 24.93 - 24.94 - 24.95 frame candidate = *fr; 24.96 24.97 // If the starting frame we were given has no codeBlob associated with 24.98 @@ -334,9 +367,11 @@ 24.99 nmethod* nm = (nmethod*) candidate.cb(); 24.100 *method_p = nm->method(); 24.101 24.102 - // If the frame isn't fully decipherable then the default 24.103 - // value for the bci is a signal that we don't have a bci. 24.104 - // If we have a decipherable frame this bci value will 24.105 + // If the frame is not decipherable, then the value of -1 24.106 + // for the BCI is used to signal that no BCI is available. 24.107 + // Furthermore, the method returns false in this case. 24.108 + // 24.109 + // If a decipherable frame is available, the BCI value will 24.110 // not be used. 24.111 24.112 *bci_p = -1; 24.113 @@ -347,9 +382,9 @@ 24.114 24.115 if (nm->is_native_method()) return true; 24.116 24.117 - // If it isn't decipherable then we have found a pc that doesn't 24.118 - // have a PCDesc that can get us a bci however we did find 24.119 - // a method 24.120 + // If the frame is not decipherable, then a PC was found 24.121 + // that does not have a PCDesc from which a BCI can be obtained. 24.122 + // Nevertheless, a Method was found. 24.123 24.124 if (!is_decipherable_compiled_frame(thread, &candidate, nm)) { 24.125 return false; 24.126 @@ -358,7 +393,7 @@ 24.127 // is_decipherable_compiled_frame may modify candidate's pc 24.128 *initial_frame_p = candidate; 24.129 24.130 - assert(nm->pc_desc_at(candidate.pc()) != NULL, "if it's decipherable then pc must be valid"); 24.131 + assert(nm->pc_desc_at(candidate.pc()) != NULL, "debug information must be available if the frame is decipherable"); 24.132 24.133 return true; 24.134 } 24.135 @@ -388,17 +423,17 @@ 24.136 24.137 frame initial_Java_frame; 24.138 Method* method; 24.139 - int bci; 24.140 + int bci = -1; // assume BCI is not available for method 24.141 + // update with correct information if available 24.142 int count; 24.143 24.144 count = 0; 24.145 assert(trace->frames != NULL, "trace->frames must be non-NULL"); 24.146 24.147 - bool fully_decipherable = find_initial_Java_frame(thd, &top_frame, &initial_Java_frame, &method, &bci); 24.148 + // Walk the stack starting from 'top_frame' and search for an initial Java frame. 24.149 + find_initial_Java_frame(thd, &top_frame, &initial_Java_frame, &method, &bci); 24.150 24.151 - // The frame might not be walkable but still recovered a method 24.152 - // (e.g. an nmethod with no scope info for the pc) 24.153 - 24.154 + // Check if a Java Method has been found. 24.155 if (method == NULL) return; 24.156 24.157 if (!method->is_valid_method()) { 24.158 @@ -406,29 +441,6 @@ 24.159 return; 24.160 } 24.161 24.162 - // We got a Java frame however it isn't fully decipherable 24.163 - // so it won't necessarily be safe to use it for the 24.164 - // initial frame in the vframe stream. 24.165 - 24.166 - if (!fully_decipherable) { 24.167 - // Take whatever method the top-frame decoder managed to scrape up. 24.168 - // We look further at the top frame only if non-safepoint 24.169 - // debugging information is available. 24.170 - count++; 24.171 - trace->num_frames = count; 24.172 - trace->frames[0].method_id = method->find_jmethod_id_or_null(); 24.173 - if (!method->is_native()) { 24.174 - trace->frames[0].lineno = bci; 24.175 - } else { 24.176 - trace->frames[0].lineno = -3; 24.177 - } 24.178 - 24.179 - if (!initial_Java_frame.safe_for_sender(thd)) return; 24.180 - 24.181 - RegisterMap map(thd, false); 24.182 - initial_Java_frame = initial_Java_frame.sender(&map); 24.183 - } 24.184 - 24.185 vframeStreamForte st(thd, initial_Java_frame, false); 24.186 24.187 for (; !st.at_end() && count < depth; st.forte_next(), count++) {
25.1 --- a/src/share/vm/runtime/globals.hpp Fri Mar 27 10:57:42 2015 +0100 25.2 +++ b/src/share/vm/runtime/globals.hpp Fri May 29 10:58:45 2015 +0200 25.3 @@ -3940,7 +3940,11 @@ 25.4 "Enable event-based tracing") \ 25.5 \ 25.6 product(bool, UseLockedTracing, false, \ 25.7 - "Use locked-tracing when doing event-based tracing") 25.8 + "Use locked-tracing when doing event-based tracing") \ 25.9 + \ 25.10 + product_pd(bool, PreserveFramePointer, \ 25.11 + "Use the FP register for holding the frame pointer " \ 25.12 + "and not as a general purpose register.") 25.13 25.14 /* 25.15 * Macros for factoring of globals
26.1 --- a/src/share/vm/runtime/sharedRuntime.cpp Fri Mar 27 10:57:42 2015 +0100 26.2 +++ b/src/share/vm/runtime/sharedRuntime.cpp Fri May 29 10:58:45 2015 +0200 26.3 @@ -1230,7 +1230,7 @@ 26.4 #endif 26.5 26.6 // JSR 292 key invariant: 26.7 - // If the resolved method is a MethodHandle invoke target the call 26.8 + // If the resolved method is a MethodHandle invoke target, the call 26.9 // site must be a MethodHandle call site, because the lambda form might tail-call 26.10 // leaving the stack in a state unknown to either caller or callee 26.11 // TODO detune for now but we might need it again
27.1 --- a/src/share/vm/runtime/vframe.hpp Fri Mar 27 10:57:42 2015 +0100 27.2 +++ b/src/share/vm/runtime/vframe.hpp Fri May 29 10:58:45 2015 +0200 27.3 @@ -389,12 +389,12 @@ 27.4 decode_offset < 0 || 27.5 decode_offset >= nm()->scopes_data_size()) { 27.6 // 6379830 AsyncGetCallTrace sometimes feeds us wild frames. 27.7 - // If we attempt to read nmethod::scopes_data at serialized_null (== 0), 27.8 - // or if we read some at other crazy offset, 27.9 - // we will decode garbage and make wild references into the heap, 27.10 - // leading to crashes in product mode. 27.11 - // (This isn't airtight, of course, since there are internal 27.12 - // offsets which are also crazy.) 27.13 + // If we read nmethod::scopes_data at serialized_null (== 0) 27.14 + // or if read some at other invalid offset, invalid values will be decoded. 27.15 + // Based on these values, invalid heap locations could be referenced 27.16 + // that could lead to crashes in product mode. 27.17 + // Therefore, do not use the decode offset if invalid, but fill the frame 27.18 + // as it were a native compiled frame (no Java-level assumptions). 27.19 #ifdef ASSERT 27.20 if (WizardMode) { 27.21 tty->print_cr("Error in fill_from_frame: pc_desc for " 27.22 @@ -514,9 +514,15 @@ 27.23 intptr_t bcx = _frame.interpreter_frame_bcx(); 27.24 int bci = method->validate_bci_from_bcx(bcx); 27.25 // 6379830 AsyncGetCallTrace sometimes feeds us wild frames. 27.26 + // AsyncGetCallTrace interrupts the VM asynchronously. As a result 27.27 + // it is possible to access an interpreter frame for which 27.28 + // no Java-level information is yet available (e.g., becasue 27.29 + // the frame was being created when the VM interrupted it). 27.30 + // In this scenario, pretend that the interpreter is at the point 27.31 + // of entering the method. 27.32 if (bci < 0) { 27.33 found_bad_method_frame(); 27.34 - bci = 0; // pretend it's on the point of entering 27.35 + bci = 0; 27.36 } 27.37 _mode = interpreted_mode; 27.38 _method = method;
