vm/inline_cache.cpp

   1 #include "master.hpp"
   2
   3 namespace factor {
   4
   5 void factor_vm::deallocate_inline_cache(cell return_address) {
   6   // Find the call target.
   7   void* old_entry_point = get_call_target(return_address);
   8   code_block* old_block = (code_block*)old_entry_point - 1;
   9
  10   // Free the old PIC since we know its unreachable
  11   if (old_block->pic_p())
  12     code->free(old_block);
  13 }
  14
  15 // Figure out what kind of type check the PIC needs based on the methods
  16 // it contains
  17 static cell determine_inline_cache_type(array* cache_entries) {
  18   for (cell i = 0; i < array_capacity(cache_entries); i += 2) {
  19     // Is it a tuple layout?
  20     if (TAG(array_nth(cache_entries, i)) == ARRAY_TYPE) {
  21       return PIC_TUPLE;
  22     }
  23   }
  24   return PIC_TAG;
  25 }
  26
  27 void factor_vm::update_pic_count(cell type) {
  28   if (type == PIC_TAG)
  29     dispatch_stats.pic_tag_count++;
  30   else
  31     dispatch_stats.pic_tuple_count++;
  32 }
  33
  34 struct inline_cache_jit : public jit {
  35   inline_cache_jit(cell generic_word, factor_vm* vm) : jit(generic_word, vm) {}
  36
  37   void emit_check_and_jump(cell ic_type, cell i, cell klass, cell method);
  38   void emit_inline_cache(fixnum index, cell generic_word_, cell methods_,
  39                          cell cache_entries_, bool tail_call_p);
  40 };
  41
  42 void inline_cache_jit::emit_check_and_jump(cell ic_type, cell i,
  43                                            cell klass, cell method) {
  44   // Class equal?
  45   cell check_type = PIC_CHECK_TAG;
  46   if (TAG(klass) != FIXNUM_TYPE)
  47       check_type = PIC_CHECK_TUPLE;
  48
  49   // The tag check can be skipped if it is the first one and we are
  50   // checking for the fixnum type which is 0. That is because the
  51   // AND instruction in the PIC_TAG template already sets the zero
  52   // flag.
  53   if (!(i == 0 && ic_type == PIC_TAG && klass == 0)) {
  54     emit_with_literal(parent->special_objects[check_type], klass);
  55   }
  56
  57   // Yes? Jump to method
  58   emit_with_literal(parent->special_objects[PIC_HIT], method);
  59 }
  60
  61 // index: 0 = top of stack, 1 = item underneath, etc
  62 // cache_entries: array of class/method pairs
  63 // Allocates memory
  64 void inline_cache_jit::emit_inline_cache(fixnum index, cell generic_word_,
  65                                          cell methods_, cell cache_entries_,
  66                                          bool tail_call_p) {
  67   data_root<word> generic_word(generic_word_, parent);
  68   data_root<array> methods(methods_, parent);
  69   data_root<array> cache_entries(cache_entries_, parent);
  70
  71   cell ic_type = determine_inline_cache_type(cache_entries.untagged());
  72   parent->update_pic_count(ic_type);
  73
  74   // Generate machine code to determine the object's class.
  75   emit_with_literal(parent->special_objects[PIC_LOAD],
  76                     tag_fixnum(-index * sizeof(cell)));
  77
  78   // Put the tag of the object, or class of the tuple in a register.
  79   emit(parent->special_objects[ic_type]);
  80
  81   // Generate machine code to check, in turn, if the class is one of the cached
  82   // entries.
  83   for (cell i = 0; i < array_capacity(cache_entries.untagged()); i += 2) {
  84     cell klass = array_nth(cache_entries.untagged(), i);
  85     cell method = array_nth(cache_entries.untagged(), i + 1);
  86
  87     emit_check_and_jump(ic_type, i, klass, method);
  88   }
  89
  90   // If none of the above conditionals tested true, then execution "falls
  91   // through" to here.
  92
  93   // A stack frame is set up, since the inline-cache-miss sub-primitive
  94   // makes a subroutine call to the VM.
  95   emit(parent->special_objects[JIT_PROLOG]);
  96
  97   // The inline-cache-miss sub-primitive call receives enough information to
  98   // reconstruct the PIC with the new entry.
  99   push(generic_word.value());
 100   push(methods.value());
 101   push(tag_fixnum(index));
 102   push(cache_entries.value());
 103
 104   emit_subprimitive(
 105       parent->special_objects[tail_call_p ? PIC_MISS_TAIL_WORD : PIC_MISS_WORD],
 106       true,  // tail_call_p
 107       true); // stack_frame_p
 108 }
 109
 110 // Allocates memory
 111 cell factor_vm::add_inline_cache_entry(cell cache_entries_, cell klass_,
 112                                        cell method_) {
 113   data_root<array> cache_entries(cache_entries_, this);
 114   data_root<object> klass(klass_, this);
 115   data_root<word> method(method_, this);
 116
 117   cell pic_size = array_capacity(cache_entries.untagged());
 118   data_root<array> new_cache_entries(
 119       reallot_array(cache_entries.untagged(), pic_size + 2), this);
 120   set_array_nth(new_cache_entries.untagged(), pic_size, klass.value());
 121   set_array_nth(new_cache_entries.untagged(), pic_size + 1, method.value());
 122   return new_cache_entries.value();
 123 }
 124
 125 void factor_vm::update_pic_transitions(cell pic_size) {
 126   if (pic_size == max_pic_size)
 127     dispatch_stats.pic_to_mega_transitions++;
 128   else if (pic_size == 0)
 129     dispatch_stats.cold_call_to_ic_transitions++;
 130   else if (pic_size == 1)
 131     dispatch_stats.ic_to_pic_transitions++;
 132 }
 133
 134 // The cache_entries parameter is empty (on cold call site) or has entries
 135 // (on cache miss). Called from assembly with the actual return address.
 136 // Compilation of the inline cache may trigger a GC, which may trigger a
 137 // compaction;
 138 // also, the block containing the return address may now be dead. Use a
 139 // code_root to take care of the details.
 140 // Allocates memory
 141 cell factor_vm::inline_cache_miss(cell return_address_) {
 142   code_root return_address(return_address_, this);
 143   bool tail_call_site = tail_call_site_p(return_address.value);
 144
 145 #ifdef PIC_DEBUG
 146   FACTOR_PRINT("Inline cache miss at "
 147                << (tail_call_site ? "tail" : "non-tail")
 148                << " call site 0x" << std::hex << return_address.value
 149                << std::dec);
 150   print_callstack();
 151 #endif
 152
 153   data_root<array> cache_entries(ctx->pop(), this);
 154   fixnum index = untag_fixnum(ctx->pop());
 155   data_root<array> methods(ctx->pop(), this);
 156   data_root<word> generic_word(ctx->pop(), this);
 157   data_root<object> object(((cell*)ctx->datastack)[-index], this);
 158
 159   cell pic_size = array_capacity(cache_entries.untagged()) / 2;
 160
 161   update_pic_transitions(pic_size);
 162
 163   cell xt = generic_word->entry_point;
 164   if (pic_size < max_pic_size) {
 165     cell klass = object_class(object.value());
 166     cell method = lookup_method(object.value(), methods.value());
 167
 168     data_root<array> new_cache_entries(
 169         add_inline_cache_entry(cache_entries.value(), klass, method), this);
 170
 171     inline_cache_jit jit(generic_word.value(), this);
 172     jit.emit_inline_cache(index, generic_word.value(), methods.value(),
 173                           new_cache_entries.value(), tail_call_site);
 174     code_block* code = jit.to_code_block(CODE_BLOCK_PIC, JIT_FRAME_SIZE);
 175     initialize_code_block(code);
 176     xt = code->entry_point();
 177   }
 178
 179   // Install the new stub.
 180   if (return_address.valid) {
 181     // Since each PIC is only referenced from a single call site,
 182     // if the old call target was a PIC, we can deallocate it immediately,
 183     // instead of leaving dead PICs around until the next GC.
 184     deallocate_inline_cache(return_address.value);
 185     set_call_target(return_address.value, xt);
 186
 187 #ifdef PIC_DEBUG
 188     FACTOR_PRINT("Updated " << (tail_call_site ? "tail" : "non-tail")
 189                  << " call site 0x" << std::hex << return_address.value << std::dec
 190                  << " with 0x" << std::hex << (cell)xt << std::dec);
 191     print_callstack();
 192 #endif
 193   }
 194
 195   return xt;
 196 }
 197
 198 // Allocates memory
 199 VM_C_API cell inline_cache_miss(cell return_address, factor_vm* parent) {
 200   return parent->inline_cache_miss(return_address);
 201 }
 202
 203 }