aoqi@0: /*
aoqi@0:  * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
aoqi@0:  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
aoqi@0:  *
aoqi@0:  * This code is free software; you can redistribute it and/or modify it
aoqi@0:  * under the terms of the GNU General Public License version 2 only, as
aoqi@0:  * published by the Free Software Foundation.
aoqi@0:  *
aoqi@0:  * This code is distributed in the hope that it will be useful, but WITHOUT
aoqi@0:  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
aoqi@0:  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
aoqi@0:  * version 2 for more details (a copy is included in the LICENSE file that
aoqi@0:  * accompanied this code).
aoqi@0:  *
aoqi@0:  * You should have received a copy of the GNU General Public License version
aoqi@0:  * 2 along with this work; if not, write to the Free Software Foundation,
aoqi@0:  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
aoqi@0:  *
aoqi@0:  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
aoqi@0:  * or visit www.oracle.com if you need additional information or have any
aoqi@0:  * questions.
aoqi@0:  *
aoqi@0:  */
aoqi@0: 
aoqi@0: #include "precompiled.hpp"
aoqi@0: #include "asm/macroAssembler.hpp"
aoqi@0: #include "asm/macroAssembler.inline.hpp"
aoqi@0: #include "asm/codeBuffer.hpp"
aoqi@0: #include "runtime/atomic.hpp"
aoqi@0: #include "runtime/atomic.inline.hpp"
aoqi@0: #include "runtime/icache.hpp"
aoqi@0: #include "runtime/os.hpp"
aoqi@0: 
aoqi@0: 
aoqi@0: // Implementation of AbstractAssembler
aoqi@0: //
aoqi@0: // The AbstractAssembler is generating code into a CodeBuffer. To make code generation faster,
aoqi@0: // the assembler keeps a copy of the code buffers boundaries & modifies them when
aoqi@0: // emitting bytes rather than using the code buffers accessor functions all the time.
aoqi@0: // The code buffer is updated via set_code_end(...) after emitting a whole instruction.
aoqi@0: 
aoqi@0: AbstractAssembler::AbstractAssembler(CodeBuffer* code) {
aoqi@0:   if (code == NULL)  return;
aoqi@0:   CodeSection* cs = code->insts();
aoqi@0:   cs->clear_mark();   // new assembler kills old mark
aoqi@0:   if (cs->start() == NULL)  {
aoqi@0:     vm_exit_out_of_memory(0, OOM_MMAP_ERROR, err_msg("CodeCache: no room for %s",
aoqi@0:                                      code->name()));
aoqi@0:   }
aoqi@0:   _code_section = cs;
aoqi@0:   _oop_recorder= code->oop_recorder();
aoqi@0:   DEBUG_ONLY( _short_branch_delta = 0; )
aoqi@0: }
aoqi@0: 
aoqi@0: void AbstractAssembler::set_code_section(CodeSection* cs) {
aoqi@0:   assert(cs->outer() == code_section()->outer(), "sanity");
aoqi@0:   assert(cs->is_allocated(), "need to pre-allocate this section");
aoqi@0:   cs->clear_mark();  // new assembly into this section kills old mark
aoqi@0:   _code_section = cs;
aoqi@0: }
aoqi@0: 
aoqi@0: // Inform CodeBuffer that incoming code and relocation will be for stubs
aoqi@0: address AbstractAssembler::start_a_stub(int required_space) {
aoqi@0:   CodeBuffer*  cb = code();
aoqi@0:   CodeSection* cs = cb->stubs();
aoqi@0:   assert(_code_section == cb->insts(), "not in insts?");
aoqi@0:   if (cs->maybe_expand_to_ensure_remaining(required_space)
aoqi@0:       && cb->blob() == NULL) {
aoqi@0:     return NULL;
aoqi@0:   }
aoqi@0:   set_code_section(cs);
aoqi@0:   return pc();
aoqi@0: }
aoqi@0: 
aoqi@0: // Inform CodeBuffer that incoming code and relocation will be code
aoqi@0: // Should not be called if start_a_stub() returned NULL
aoqi@0: void AbstractAssembler::end_a_stub() {
aoqi@0:   assert(_code_section == code()->stubs(), "not in stubs?");
aoqi@0:   set_code_section(code()->insts());
aoqi@0: }
aoqi@0: 
aoqi@0: // Inform CodeBuffer that incoming code and relocation will be for stubs
aoqi@0: address AbstractAssembler::start_a_const(int required_space, int required_align) {
aoqi@0:   CodeBuffer*  cb = code();
aoqi@0:   CodeSection* cs = cb->consts();
aoqi@0:   assert(_code_section == cb->insts() || _code_section == cb->stubs(), "not in insts/stubs?");
aoqi@0:   address end = cs->end();
aoqi@0:   int pad = -(intptr_t)end & (required_align-1);
aoqi@0:   if (cs->maybe_expand_to_ensure_remaining(pad + required_space)) {
aoqi@0:     if (cb->blob() == NULL)  return NULL;
aoqi@0:     end = cs->end();  // refresh pointer
aoqi@0:   }
aoqi@0:   if (pad > 0) {
aoqi@0:     while (--pad >= 0) { *end++ = 0; }
aoqi@0:     cs->set_end(end);
aoqi@0:   }
aoqi@0:   set_code_section(cs);
aoqi@0:   return end;
aoqi@0: }
aoqi@0: 
aoqi@0: // Inform CodeBuffer that incoming code and relocation will be code
aoqi@0: // in section cs (insts or stubs).
aoqi@0: void AbstractAssembler::end_a_const(CodeSection* cs) {
aoqi@0:   assert(_code_section == code()->consts(), "not in consts?");
aoqi@0:   set_code_section(cs);
aoqi@0: }
aoqi@0: 
aoqi@0: void AbstractAssembler::flush() {
aoqi@0:   ICache::invalidate_range(addr_at(0), offset());
aoqi@0: }
aoqi@0: 
aoqi@0: void AbstractAssembler::bind(Label& L) {
aoqi@0:   if (L.is_bound()) {
aoqi@0:     // Assembler can bind a label more than once to the same place.
aoqi@0:     guarantee(L.loc() == locator(), "attempt to redefine label");
aoqi@0:     return;
aoqi@0:   }
aoqi@0:   L.bind_loc(locator());
aoqi@0:   L.patch_instructions((MacroAssembler*)this);
aoqi@0: }
aoqi@0: 
aoqi@0: void AbstractAssembler::generate_stack_overflow_check( int frame_size_in_bytes) {
aoqi@0:   if (UseStackBanging) {
aoqi@0:     // Each code entry causes one stack bang n pages down the stack where n
aoqi@0:     // is configurable by StackShadowPages.  The setting depends on the maximum
aoqi@0:     // depth of VM call stack or native before going back into java code,
aoqi@0:     // since only java code can raise a stack overflow exception using the
aoqi@0:     // stack banging mechanism.  The VM and native code does not detect stack
aoqi@0:     // overflow.
aoqi@0:     // The code in JavaCalls::call() checks that there is at least n pages
aoqi@0:     // available, so all entry code needs to do is bang once for the end of
aoqi@0:     // this shadow zone.
aoqi@0:     // The entry code may need to bang additional pages if the framesize
aoqi@0:     // is greater than a page.
aoqi@0: 
aoqi@0:     const int page_size = os::vm_page_size();
aoqi@0:     int bang_end = StackShadowPages*page_size;
aoqi@0: 
aoqi@0:     // This is how far the previous frame's stack banging extended.
aoqi@0:     const int bang_end_safe = bang_end;
aoqi@0: 
aoqi@0:     if (frame_size_in_bytes > page_size) {
aoqi@0:       bang_end += frame_size_in_bytes;
aoqi@0:     }
aoqi@0: 
aoqi@0:     int bang_offset = bang_end_safe;
aoqi@0:     while (bang_offset <= bang_end) {
aoqi@0:       // Need at least one stack bang at end of shadow zone.
aoqi@0:       bang_stack_with_offset(bang_offset);
aoqi@0:       bang_offset += page_size;
aoqi@0:     }
aoqi@0:   } // end (UseStackBanging)
aoqi@0: }
aoqi@0: 
aoqi@0: void Label::add_patch_at(CodeBuffer* cb, int branch_loc) {
aoqi@0:   assert(_loc == -1, "Label is unbound");
aoqi@0:   if (_patch_index < PatchCacheSize) {
aoqi@0:     _patches[_patch_index] = branch_loc;
aoqi@0:   } else {
aoqi@0:     if (_patch_overflow == NULL) {
aoqi@0:       _patch_overflow = cb->create_patch_overflow();
aoqi@0:     }
aoqi@0:     _patch_overflow->push(branch_loc);
aoqi@0:   }
aoqi@0:   ++_patch_index;
aoqi@0: }
aoqi@0: 
aoqi@0: void Label::patch_instructions(MacroAssembler* masm) {
aoqi@0:   assert(is_bound(), "Label is bound");
aoqi@0:   CodeBuffer* cb = masm->code();
aoqi@0:   int target_sect = CodeBuffer::locator_sect(loc());
aoqi@0:   address target = cb->locator_address(loc());
aoqi@0:   while (_patch_index > 0) {
aoqi@0:     --_patch_index;
aoqi@0:     int branch_loc;
aoqi@0:     if (_patch_index >= PatchCacheSize) {
aoqi@0:       branch_loc = _patch_overflow->pop();
aoqi@0:     } else {
aoqi@0:       branch_loc = _patches[_patch_index];
aoqi@0:     }
aoqi@0:     int branch_sect = CodeBuffer::locator_sect(branch_loc);
aoqi@0:     address branch = cb->locator_address(branch_loc);
aoqi@0:     if (branch_sect == CodeBuffer::SECT_CONSTS) {
aoqi@0:       // The thing to patch is a constant word.
aoqi@0:       *(address*)branch = target;
aoqi@0:       continue;
aoqi@0:     }
aoqi@0: 
aoqi@0: #ifdef ASSERT
aoqi@0:     // Cross-section branches only work if the
aoqi@0:     // intermediate section boundaries are frozen.
aoqi@0:     if (target_sect != branch_sect) {
aoqi@0:       for (int n = MIN2(target_sect, branch_sect),
aoqi@0:                nlimit = (target_sect + branch_sect) - n;
aoqi@0:            n < nlimit; n++) {
aoqi@0:         CodeSection* cs = cb->code_section(n);
aoqi@0:         assert(cs->is_frozen(), "cross-section branch needs stable offsets");
aoqi@0:       }
aoqi@0:     }
aoqi@0: #endif //ASSERT
aoqi@0: 
aoqi@0:     // Push the target offset into the branch instruction.
aoqi@0:     masm->pd_patch_instruction(branch, target);
aoqi@0:   }
aoqi@0: }
aoqi@0: 
aoqi@0: struct DelayedConstant {
aoqi@0:   typedef void (*value_fn_t)();
aoqi@0:   BasicType type;
aoqi@0:   intptr_t value;
aoqi@0:   value_fn_t value_fn;
aoqi@0:   // This limit of 20 is generous for initial uses.
aoqi@0:   // The limit needs to be large enough to store the field offsets
aoqi@0:   // into classes which do not have statically fixed layouts.
aoqi@0:   // (Initial use is for method handle object offsets.)
aoqi@0:   // Look for uses of "delayed_value" in the source code
aoqi@0:   // and make sure this number is generous enough to handle all of them.
aoqi@0:   enum { DC_LIMIT = 20 };
aoqi@0:   static DelayedConstant delayed_constants[DC_LIMIT];
aoqi@0:   static DelayedConstant* add(BasicType type, value_fn_t value_fn);
aoqi@0:   bool match(BasicType t, value_fn_t cfn) {
aoqi@0:     return type == t && value_fn == cfn;
aoqi@0:   }
aoqi@0:   static void update_all();
aoqi@0: };
aoqi@0: 
aoqi@0: DelayedConstant DelayedConstant::delayed_constants[DC_LIMIT];
aoqi@0: // Default C structure initialization rules have the following effect here:
aoqi@0: // = { { (BasicType)0, (intptr_t)NULL }, ... };
aoqi@0: 
aoqi@0: DelayedConstant* DelayedConstant::add(BasicType type,
aoqi@0:                                       DelayedConstant::value_fn_t cfn) {
aoqi@0:   for (int i = 0; i < DC_LIMIT; i++) {
aoqi@0:     DelayedConstant* dcon = &delayed_constants[i];
aoqi@0:     if (dcon->match(type, cfn))
aoqi@0:       return dcon;
aoqi@0:     if (dcon->value_fn == NULL) {
aoqi@0:       // (cmpxchg not because this is multi-threaded but because I'm paranoid)
aoqi@0:       if (Atomic::cmpxchg_ptr(CAST_FROM_FN_PTR(void*, cfn), &dcon->value_fn, NULL) == NULL) {
aoqi@0:         dcon->type = type;
aoqi@0:         return dcon;
aoqi@0:       }
aoqi@0:     }
aoqi@0:   }
aoqi@0:   // If this assert is hit (in pre-integration testing!) then re-evaluate
aoqi@0:   // the comment on the definition of DC_LIMIT.
aoqi@0:   guarantee(false, "too many delayed constants");
aoqi@0:   return NULL;
aoqi@0: }
aoqi@0: 
aoqi@0: void DelayedConstant::update_all() {
aoqi@0:   for (int i = 0; i < DC_LIMIT; i++) {
aoqi@0:     DelayedConstant* dcon = &delayed_constants[i];
aoqi@0:     if (dcon->value_fn != NULL && dcon->value == 0) {
aoqi@0:       typedef int     (*int_fn_t)();
aoqi@0:       typedef address (*address_fn_t)();
aoqi@0:       switch (dcon->type) {
aoqi@0:       case T_INT:     dcon->value = (intptr_t) ((int_fn_t)    dcon->value_fn)(); break;
aoqi@0:       case T_ADDRESS: dcon->value = (intptr_t) ((address_fn_t)dcon->value_fn)(); break;
aoqi@0:       }
aoqi@0:     }
aoqi@0:   }
aoqi@0: }
aoqi@0: 
aoqi@0: RegisterOrConstant AbstractAssembler::delayed_value(int(*value_fn)(), Register tmp, int offset) {
aoqi@0:   intptr_t val = (intptr_t) (*value_fn)();
aoqi@0:   if (val != 0)  return val + offset;
aoqi@0:   return delayed_value_impl(delayed_value_addr(value_fn), tmp, offset);
aoqi@0: }
aoqi@0: RegisterOrConstant AbstractAssembler::delayed_value(address(*value_fn)(), Register tmp, int offset) {
aoqi@0:   intptr_t val = (intptr_t) (*value_fn)();
aoqi@0:   if (val != 0)  return val + offset;
aoqi@0:   return delayed_value_impl(delayed_value_addr(value_fn), tmp, offset);
aoqi@0: }
aoqi@0: intptr_t* AbstractAssembler::delayed_value_addr(int(*value_fn)()) {
aoqi@0:   DelayedConstant* dcon = DelayedConstant::add(T_INT, (DelayedConstant::value_fn_t) value_fn);
aoqi@0:   return &dcon->value;
aoqi@0: }
aoqi@0: intptr_t* AbstractAssembler::delayed_value_addr(address(*value_fn)()) {
aoqi@0:   DelayedConstant* dcon = DelayedConstant::add(T_ADDRESS, (DelayedConstant::value_fn_t) value_fn);
aoqi@0:   return &dcon->value;
aoqi@0: }
aoqi@0: void AbstractAssembler::update_delayed_values() {
aoqi@0:   DelayedConstant::update_all();
aoqi@0: }
aoqi@0: 
aoqi@0: void AbstractAssembler::block_comment(const char* comment) {
aoqi@0:   if (sect() == CodeBuffer::SECT_INSTS) {
aoqi@0:     code_section()->outer()->block_comment(offset(), comment);
aoqi@0:   }
aoqi@0: }
aoqi@0: 
aoqi@0: const char* AbstractAssembler::code_string(const char* str) {
aoqi@0:   if (sect() == CodeBuffer::SECT_INSTS || sect() == CodeBuffer::SECT_STUBS) {
aoqi@0:     return code_section()->outer()->code_string(str);
aoqi@0:   }
aoqi@0:   return NULL;
aoqi@0: }
aoqi@0: 
aoqi@0: bool MacroAssembler::needs_explicit_null_check(intptr_t offset) {
aoqi@0:   // Exception handler checks the nmethod's implicit null checks table
aoqi@0:   // only when this method returns false.
aoqi@0: #ifdef _LP64
aoqi@0:   if (UseCompressedOops && Universe::narrow_oop_base() != NULL) {
aoqi@0:     assert (Universe::heap() != NULL, "java heap should be initialized");
aoqi@0:     // The first page after heap_base is unmapped and
aoqi@0:     // the 'offset' is equal to [heap_base + offset] for
aoqi@0:     // narrow oop implicit null checks.
aoqi@0:     uintptr_t base = (uintptr_t)Universe::narrow_oop_base();
aoqi@0:     if ((uintptr_t)offset >= base) {
aoqi@0:       // Normalize offset for the next check.
aoqi@0:       offset = (intptr_t)(pointer_delta((void*)offset, (void*)base, 1));
aoqi@0:     }
aoqi@0:   }
aoqi@0: #endif
aoqi@0:   return offset < 0 || os::vm_page_size() <= offset;
aoqi@0: }