1.1 --- a/src/cpu/x86/vm/x86_32.ad Wed Feb 15 16:29:40 2012 -0800
1.2 +++ b/src/cpu/x86/vm/x86_32.ad Wed Feb 15 21:37:49 2012 -0800
1.3 @@ -550,118 +550,66 @@
1.4
1.5 //=============================================================================
1.6 #ifndef PRODUCT
1.7 -void MachPrologNode::format( PhaseRegAlloc *ra_, outputStream* st ) const {
1.8 +void MachPrologNode::format(PhaseRegAlloc* ra_, outputStream* st) const {
1.9 Compile* C = ra_->C;
1.10 - if( C->in_24_bit_fp_mode() ) {
1.11 - st->print("FLDCW 24 bit fpu control word");
1.12 - st->print_cr(""); st->print("\t");
1.13 - }
1.14
1.15 int framesize = C->frame_slots() << LogBytesPerInt;
1.16 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
1.17 - // Remove two words for return addr and rbp,
1.18 - framesize -= 2*wordSize;
1.19 -
1.20 - // Calls to C2R adapters often do not accept exceptional returns.
1.21 - // We require that their callers must bang for them. But be careful, because
1.22 - // some VM calls (such as call site linkage) can use several kilobytes of
1.23 - // stack. But the stack safety zone should account for that.
1.24 - // See bugs 4446381, 4468289, 4497237.
1.25 + // Remove wordSize for return addr which is already pushed.
1.26 + framesize -= wordSize;
1.27 +
1.28 if (C->need_stack_bang(framesize)) {
1.29 - st->print_cr("# stack bang"); st->print("\t");
1.30 - }
1.31 - st->print_cr("PUSHL EBP"); st->print("\t");
1.32 -
1.33 - if( VerifyStackAtCalls ) { // Majik cookie to verify stack depth
1.34 - st->print("PUSH 0xBADB100D\t# Majik cookie for stack depth check");
1.35 - st->print_cr(""); st->print("\t");
1.36 framesize -= wordSize;
1.37 - }
1.38 -
1.39 - if ((C->in_24_bit_fp_mode() || VerifyStackAtCalls ) && framesize < 128 ) {
1.40 + st->print("# stack bang");
1.41 + st->print("\n\t");
1.42 + st->print("PUSH EBP\t# Save EBP");
1.43 if (framesize) {
1.44 - st->print("SUB ESP,%d\t# Create frame",framesize);
1.45 + st->print("\n\t");
1.46 + st->print("SUB ESP, #%d\t# Create frame",framesize);
1.47 }
1.48 } else {
1.49 - st->print("SUB ESP,%d\t# Create frame",framesize);
1.50 + st->print("SUB ESP, #%d\t# Create frame",framesize);
1.51 + st->print("\n\t");
1.52 + framesize -= wordSize;
1.53 + st->print("MOV [ESP + #%d], EBP\t# Save EBP",framesize);
1.54 }
1.55 +
1.56 + if (VerifyStackAtCalls) {
1.57 + st->print("\n\t");
1.58 + framesize -= wordSize;
1.59 + st->print("MOV [ESP + #%d], 0xBADB100D\t# Majik cookie for stack depth check",framesize);
1.60 + }
1.61 +
1.62 + if( C->in_24_bit_fp_mode() ) {
1.63 + st->print("\n\t");
1.64 + st->print("FLDCW \t# load 24 bit fpu control word");
1.65 + }
1.66 + if (UseSSE >= 2 && VerifyFPU) {
1.67 + st->print("\n\t");
1.68 + st->print("# verify FPU stack (must be clean on entry)");
1.69 + }
1.70 +
1.71 +#ifdef ASSERT
1.72 + if (VerifyStackAtCalls) {
1.73 + st->print("\n\t");
1.74 + st->print("# stack alignment check");
1.75 + }
1.76 +#endif
1.77 + st->cr();
1.78 }
1.79 #endif
1.80
1.81
1.82 void MachPrologNode::emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const {
1.83 Compile* C = ra_->C;
1.84 -
1.85 - if (UseSSE >= 2 && VerifyFPU) {
1.86 - MacroAssembler masm(&cbuf);
1.87 - masm.verify_FPU(0, "FPU stack must be clean on entry");
1.88 - }
1.89 -
1.90 - // WARNING: Initial instruction MUST be 5 bytes or longer so that
1.91 - // NativeJump::patch_verified_entry will be able to patch out the entry
1.92 - // code safely. The fldcw is ok at 6 bytes, the push to verify stack
1.93 - // depth is ok at 5 bytes, the frame allocation can be either 3 or
1.94 - // 6 bytes. So if we don't do the fldcw or the push then we must
1.95 - // use the 6 byte frame allocation even if we have no frame. :-(
1.96 - // If method sets FPU control word do it now
1.97 - if( C->in_24_bit_fp_mode() ) {
1.98 - MacroAssembler masm(&cbuf);
1.99 - masm.fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24()));
1.100 - }
1.101 + MacroAssembler _masm(&cbuf);
1.102
1.103 int framesize = C->frame_slots() << LogBytesPerInt;
1.104 - assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
1.105 - // Remove two words for return addr and rbp,
1.106 - framesize -= 2*wordSize;
1.107 -
1.108 - // Calls to C2R adapters often do not accept exceptional returns.
1.109 - // We require that their callers must bang for them. But be careful, because
1.110 - // some VM calls (such as call site linkage) can use several kilobytes of
1.111 - // stack. But the stack safety zone should account for that.
1.112 - // See bugs 4446381, 4468289, 4497237.
1.113 - if (C->need_stack_bang(framesize)) {
1.114 - MacroAssembler masm(&cbuf);
1.115 - masm.generate_stack_overflow_check(framesize);
1.116 - }
1.117 -
1.118 - // We always push rbp, so that on return to interpreter rbp, will be
1.119 - // restored correctly and we can correct the stack.
1.120 - emit_opcode(cbuf, 0x50 | EBP_enc);
1.121 -
1.122 - if( VerifyStackAtCalls ) { // Majik cookie to verify stack depth
1.123 - emit_opcode(cbuf, 0x68); // push 0xbadb100d
1.124 - emit_d32(cbuf, 0xbadb100d);
1.125 - framesize -= wordSize;
1.126 - }
1.127 -
1.128 - if ((C->in_24_bit_fp_mode() || VerifyStackAtCalls ) && framesize < 128 ) {
1.129 - if (framesize) {
1.130 - emit_opcode(cbuf, 0x83); // sub SP,#framesize
1.131 - emit_rm(cbuf, 0x3, 0x05, ESP_enc);
1.132 - emit_d8(cbuf, framesize);
1.133 - }
1.134 - } else {
1.135 - emit_opcode(cbuf, 0x81); // sub SP,#framesize
1.136 - emit_rm(cbuf, 0x3, 0x05, ESP_enc);
1.137 - emit_d32(cbuf, framesize);
1.138 - }
1.139 +
1.140 + __ verified_entry(framesize, C->need_stack_bang(framesize), C->in_24_bit_fp_mode());
1.141 +
1.142 C->set_frame_complete(cbuf.insts_size());
1.143
1.144 -#ifdef ASSERT
1.145 - if (VerifyStackAtCalls) {
1.146 - Label L;
1.147 - MacroAssembler masm(&cbuf);
1.148 - masm.push(rax);
1.149 - masm.mov(rax, rsp);
1.150 - masm.andptr(rax, StackAlignmentInBytes-1);
1.151 - masm.cmpptr(rax, StackAlignmentInBytes-wordSize);
1.152 - masm.pop(rax);
1.153 - masm.jcc(Assembler::equal, L);
1.154 - masm.stop("Stack is not properly aligned!");
1.155 - masm.bind(L);
1.156 - }
1.157 -#endif
1.158 -
1.159 if (C->has_mach_constant_base_node()) {
1.160 // NOTE: We set the table base offset here because users might be
1.161 // emitted before MachConstantBaseNode.
