src/share/vm/asm/assembler.cpp

Fri, 30 Nov 2012 11:44:05 -0800

author
twisti
date
Fri, 30 Nov 2012 11:44:05 -0800
changeset 4317
6ab62ad83507
parent 4316
1acccb7c0b01
child 4318
cd3d6a6b95d9
permissions
-rw-r--r--

8003195: AbstractAssembler should not store code pointers but use the CodeSection directly
Reviewed-by: twisti, kvn
Contributed-by: Bharadwaj Yadavalli <bharadwaj.yadavalli@oracle.com>

     1 /*
     2  * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved.
     3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
     4  *
     5  * This code is free software; you can redistribute it and/or modify it
     6  * under the terms of the GNU General Public License version 2 only, as
     7  * published by the Free Software Foundation.
     8  *
     9  * This code is distributed in the hope that it will be useful, but WITHOUT
    10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
    11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
    12  * version 2 for more details (a copy is included in the LICENSE file that
    13  * accompanied this code).
    14  *
    15  * You should have received a copy of the GNU General Public License version
    16  * 2 along with this work; if not, write to the Free Software Foundation,
    17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
    18  *
    19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
    20  * or visit www.oracle.com if you need additional information or have any
    21  * questions.
    22  *
    23  */
    25 #include "precompiled.hpp"
    26 #include "asm/assembler.hpp"
    27 #include "asm/assembler.inline.hpp"
    28 #include "asm/codeBuffer.hpp"
    29 #include "runtime/icache.hpp"
    30 #include "runtime/os.hpp"
    31 #ifdef TARGET_ARCH_x86
    32 # include "assembler_x86.inline.hpp"
    33 #endif
    34 #ifdef TARGET_ARCH_sparc
    35 # include "assembler_sparc.inline.hpp"
    36 #endif
    37 #ifdef TARGET_ARCH_zero
    38 # include "assembler_zero.inline.hpp"
    39 #endif
    40 #ifdef TARGET_ARCH_arm
    41 # include "assembler_arm.inline.hpp"
    42 #endif
    43 #ifdef TARGET_ARCH_ppc
    44 # include "assembler_ppc.inline.hpp"
    45 #endif
    48 // Implementation of AbstractAssembler
    49 //
    50 // The AbstractAssembler is generating code into a CodeBuffer. To make code generation faster,
    51 // the assembler keeps a copy of the code buffers boundaries & modifies them when
    52 // emitting bytes rather than using the code buffers accessor functions all the time.
    53 // The code buffer is updated via set_code_end(...) after emitting a whole instruction.
    55 AbstractAssembler::AbstractAssembler(CodeBuffer* code) {
    56   if (code == NULL)  return;
    57   CodeSection* cs = code->insts();
    58   cs->clear_mark();   // new assembler kills old mark
    59   if (cs->start() == NULL)  {
    60     vm_exit_out_of_memory(0, err_msg("CodeCache: no room for %s",
    61                                      code->name()));
    62   }
    63   _code_section = cs;
    64   _oop_recorder= code->oop_recorder();
    65   DEBUG_ONLY( _short_branch_delta = 0; )
    66 }
    68 void AbstractAssembler::set_code_section(CodeSection* cs) {
    69   assert(cs->outer() == code_section()->outer(), "sanity");
    70   assert(cs->is_allocated(), "need to pre-allocate this section");
    71   cs->clear_mark();  // new assembly into this section kills old mark
    72   _code_section = cs;
    73 }
    75 // Inform CodeBuffer that incoming code and relocation will be for stubs
    76 address AbstractAssembler::start_a_stub(int required_space) {
    77   CodeBuffer*  cb = code();
    78   CodeSection* cs = cb->stubs();
    79   assert(_code_section == cb->insts(), "not in insts?");
    80   if (cs->maybe_expand_to_ensure_remaining(required_space)
    81       && cb->blob() == NULL) {
    82     return NULL;
    83   }
    84   set_code_section(cs);
    85   return pc();
    86 }
    88 // Inform CodeBuffer that incoming code and relocation will be code
    89 // Should not be called if start_a_stub() returned NULL
    90 void AbstractAssembler::end_a_stub() {
    91   assert(_code_section == code()->stubs(), "not in stubs?");
    92   set_code_section(code()->insts());
    93 }
    95 // Inform CodeBuffer that incoming code and relocation will be for stubs
    96 address AbstractAssembler::start_a_const(int required_space, int required_align) {
    97   CodeBuffer*  cb = code();
    98   CodeSection* cs = cb->consts();
    99   assert(_code_section == cb->insts() || _code_section == cb->stubs(), "not in insts/stubs?");
   100   address end = cs->end();
   101   int pad = -(intptr_t)end & (required_align-1);
   102   if (cs->maybe_expand_to_ensure_remaining(pad + required_space)) {
   103     if (cb->blob() == NULL)  return NULL;
   104     end = cs->end();  // refresh pointer
   105   }
   106   if (pad > 0) {
   107     while (--pad >= 0) { *end++ = 0; }
   108     cs->set_end(end);
   109   }
   110   set_code_section(cs);
   111   return end;
   112 }
   114 // Inform CodeBuffer that incoming code and relocation will be code
   115 // in section cs (insts or stubs).
   116 void AbstractAssembler::end_a_const(CodeSection* cs) {
   117   assert(_code_section == code()->consts(), "not in consts?");
   118   set_code_section(cs);
   119 }
   121 void AbstractAssembler::flush() {
   122   ICache::invalidate_range(addr_at(0), offset());
   123 }
   126 void AbstractAssembler::a_byte(int x) {
   127   emit_byte(x);
   128 }
   131 void AbstractAssembler::a_long(jint x) {
   132   emit_long(x);
   133 }
   135 // Labels refer to positions in the (to be) generated code.  There are bound
   136 // and unbound
   137 //
   138 // Bound labels refer to known positions in the already generated code.
   139 // offset() is the position the label refers to.
   140 //
   141 // Unbound labels refer to unknown positions in the code to be generated; it
   142 // may contain a list of unresolved displacements that refer to it
   143 #ifndef PRODUCT
   144 void AbstractAssembler::print(Label& L) {
   145   if (L.is_bound()) {
   146     tty->print_cr("bound label to %d|%d", L.loc_pos(), L.loc_sect());
   147   } else if (L.is_unbound()) {
   148     L.print_instructions((MacroAssembler*)this);
   149   } else {
   150     tty->print_cr("label in inconsistent state (loc = %d)", L.loc());
   151   }
   152 }
   153 #endif // PRODUCT
   156 void AbstractAssembler::bind(Label& L) {
   157   if (L.is_bound()) {
   158     // Assembler can bind a label more than once to the same place.
   159     guarantee(L.loc() == locator(), "attempt to redefine label");
   160     return;
   161   }
   162   L.bind_loc(locator());
   163   L.patch_instructions((MacroAssembler*)this);
   164 }
   166 void AbstractAssembler::generate_stack_overflow_check( int frame_size_in_bytes) {
   167   if (UseStackBanging) {
   168     // Each code entry causes one stack bang n pages down the stack where n
   169     // is configurable by StackBangPages.  The setting depends on the maximum
   170     // depth of VM call stack or native before going back into java code,
   171     // since only java code can raise a stack overflow exception using the
   172     // stack banging mechanism.  The VM and native code does not detect stack
   173     // overflow.
   174     // The code in JavaCalls::call() checks that there is at least n pages
   175     // available, so all entry code needs to do is bang once for the end of
   176     // this shadow zone.
   177     // The entry code may need to bang additional pages if the framesize
   178     // is greater than a page.
   180     const int page_size = os::vm_page_size();
   181     int bang_end = StackShadowPages*page_size;
   183     // This is how far the previous frame's stack banging extended.
   184     const int bang_end_safe = bang_end;
   186     if (frame_size_in_bytes > page_size) {
   187       bang_end += frame_size_in_bytes;
   188     }
   190     int bang_offset = bang_end_safe;
   191     while (bang_offset <= bang_end) {
   192       // Need at least one stack bang at end of shadow zone.
   193       bang_stack_with_offset(bang_offset);
   194       bang_offset += page_size;
   195     }
   196   } // end (UseStackBanging)
   197 }
   199 void Label::add_patch_at(CodeBuffer* cb, int branch_loc) {
   200   assert(_loc == -1, "Label is unbound");
   201   if (_patch_index < PatchCacheSize) {
   202     _patches[_patch_index] = branch_loc;
   203   } else {
   204     if (_patch_overflow == NULL) {
   205       _patch_overflow = cb->create_patch_overflow();
   206     }
   207     _patch_overflow->push(branch_loc);
   208   }
   209   ++_patch_index;
   210 }
   212 void Label::patch_instructions(MacroAssembler* masm) {
   213   assert(is_bound(), "Label is bound");
   214   CodeBuffer* cb = masm->code();
   215   int target_sect = CodeBuffer::locator_sect(loc());
   216   address target = cb->locator_address(loc());
   217   while (_patch_index > 0) {
   218     --_patch_index;
   219     int branch_loc;
   220     if (_patch_index >= PatchCacheSize) {
   221       branch_loc = _patch_overflow->pop();
   222     } else {
   223       branch_loc = _patches[_patch_index];
   224     }
   225     int branch_sect = CodeBuffer::locator_sect(branch_loc);
   226     address branch = cb->locator_address(branch_loc);
   227     if (branch_sect == CodeBuffer::SECT_CONSTS) {
   228       // The thing to patch is a constant word.
   229       *(address*)branch = target;
   230       continue;
   231     }
   233 #ifdef ASSERT
   234     // Cross-section branches only work if the
   235     // intermediate section boundaries are frozen.
   236     if (target_sect != branch_sect) {
   237       for (int n = MIN2(target_sect, branch_sect),
   238                nlimit = (target_sect + branch_sect) - n;
   239            n < nlimit; n++) {
   240         CodeSection* cs = cb->code_section(n);
   241         assert(cs->is_frozen(), "cross-section branch needs stable offsets");
   242       }
   243     }
   244 #endif //ASSERT
   246     // Push the target offset into the branch instruction.
   247     masm->pd_patch_instruction(branch, target);
   248   }
   249 }
   251 struct DelayedConstant {
   252   typedef void (*value_fn_t)();
   253   BasicType type;
   254   intptr_t value;
   255   value_fn_t value_fn;
   256   // This limit of 20 is generous for initial uses.
   257   // The limit needs to be large enough to store the field offsets
   258   // into classes which do not have statically fixed layouts.
   259   // (Initial use is for method handle object offsets.)
   260   // Look for uses of "delayed_value" in the source code
   261   // and make sure this number is generous enough to handle all of them.
   262   enum { DC_LIMIT = 20 };
   263   static DelayedConstant delayed_constants[DC_LIMIT];
   264   static DelayedConstant* add(BasicType type, value_fn_t value_fn);
   265   bool match(BasicType t, value_fn_t cfn) {
   266     return type == t && value_fn == cfn;
   267   }
   268   static void update_all();
   269 };
   271 DelayedConstant DelayedConstant::delayed_constants[DC_LIMIT];
   272 // Default C structure initialization rules have the following effect here:
   273 // = { { (BasicType)0, (intptr_t)NULL }, ... };
   275 DelayedConstant* DelayedConstant::add(BasicType type,
   276                                       DelayedConstant::value_fn_t cfn) {
   277   for (int i = 0; i < DC_LIMIT; i++) {
   278     DelayedConstant* dcon = &delayed_constants[i];
   279     if (dcon->match(type, cfn))
   280       return dcon;
   281     if (dcon->value_fn == NULL) {
   282       // (cmpxchg not because this is multi-threaded but because I'm paranoid)
   283       if (Atomic::cmpxchg_ptr(CAST_FROM_FN_PTR(void*, cfn), &dcon->value_fn, NULL) == NULL) {
   284         dcon->type = type;
   285         return dcon;
   286       }
   287     }
   288   }
   289   // If this assert is hit (in pre-integration testing!) then re-evaluate
   290   // the comment on the definition of DC_LIMIT.
   291   guarantee(false, "too many delayed constants");
   292   return NULL;
   293 }
   295 void DelayedConstant::update_all() {
   296   for (int i = 0; i < DC_LIMIT; i++) {
   297     DelayedConstant* dcon = &delayed_constants[i];
   298     if (dcon->value_fn != NULL && dcon->value == 0) {
   299       typedef int     (*int_fn_t)();
   300       typedef address (*address_fn_t)();
   301       switch (dcon->type) {
   302       case T_INT:     dcon->value = (intptr_t) ((int_fn_t)    dcon->value_fn)(); break;
   303       case T_ADDRESS: dcon->value = (intptr_t) ((address_fn_t)dcon->value_fn)(); break;
   304       }
   305     }
   306   }
   307 }
   309 RegisterOrConstant AbstractAssembler::delayed_value(int(*value_fn)(), Register tmp, int offset) {
   310   intptr_t val = (intptr_t) (*value_fn)();
   311   if (val != 0)  return val + offset;
   312   return delayed_value_impl(delayed_value_addr(value_fn), tmp, offset);
   313 }
   314 RegisterOrConstant AbstractAssembler::delayed_value(address(*value_fn)(), Register tmp, int offset) {
   315   intptr_t val = (intptr_t) (*value_fn)();
   316   if (val != 0)  return val + offset;
   317   return delayed_value_impl(delayed_value_addr(value_fn), tmp, offset);
   318 }
   319 intptr_t* AbstractAssembler::delayed_value_addr(int(*value_fn)()) {
   320   DelayedConstant* dcon = DelayedConstant::add(T_INT, (DelayedConstant::value_fn_t) value_fn);
   321   return &dcon->value;
   322 }
   323 intptr_t* AbstractAssembler::delayed_value_addr(address(*value_fn)()) {
   324   DelayedConstant* dcon = DelayedConstant::add(T_ADDRESS, (DelayedConstant::value_fn_t) value_fn);
   325   return &dcon->value;
   326 }
   327 void AbstractAssembler::update_delayed_values() {
   328   DelayedConstant::update_all();
   329 }
   334 void AbstractAssembler::block_comment(const char* comment) {
   335   if (sect() == CodeBuffer::SECT_INSTS) {
   336     code_section()->outer()->block_comment(offset(), comment);
   337   }
   338 }
   340 bool MacroAssembler::needs_explicit_null_check(intptr_t offset) {
   341   // Exception handler checks the nmethod's implicit null checks table
   342   // only when this method returns false.
   343 #ifdef _LP64
   344   if (UseCompressedOops && Universe::narrow_oop_base() != NULL) {
   345     assert (Universe::heap() != NULL, "java heap should be initialized");
   346     // The first page after heap_base is unmapped and
   347     // the 'offset' is equal to [heap_base + offset] for
   348     // narrow oop implicit null checks.
   349     uintptr_t base = (uintptr_t)Universe::narrow_oop_base();
   350     if ((uintptr_t)offset >= base) {
   351       // Normalize offset for the next check.
   352       offset = (intptr_t)(pointer_delta((void*)offset, (void*)base, 1));
   353     }
   354   }
   355 #endif
   356   return offset < 0 || os::vm_page_size() <= offset;
   357 }
   359 #ifndef PRODUCT
   360 void Label::print_instructions(MacroAssembler* masm) const {
   361   CodeBuffer* cb = masm->code();
   362   for (int i = 0; i < _patch_index; ++i) {
   363     int branch_loc;
   364     if (i >= PatchCacheSize) {
   365       branch_loc = _patch_overflow->at(i - PatchCacheSize);
   366     } else {
   367       branch_loc = _patches[i];
   368     }
   369     int branch_pos  = CodeBuffer::locator_pos(branch_loc);
   370     int branch_sect = CodeBuffer::locator_sect(branch_loc);
   371     address branch = cb->locator_address(branch_loc);
   372     tty->print_cr("unbound label");
   373     tty->print("@ %d|%d ", branch_pos, branch_sect);
   374     if (branch_sect == CodeBuffer::SECT_CONSTS) {
   375       tty->print_cr(PTR_FORMAT, *(address*)branch);
   376       continue;
   377     }
   378     masm->pd_print_patched_instruction(branch);
   379     tty->cr();
   380   }
   381 }
   382 #endif // ndef PRODUCT

mercurial