// The cache_entries parameter is empty (on cold call site) or has entries
// (on cache miss). Called from assembly with the actual return address.
// Compilation of the inline cache may trigger a GC, which may trigger a
// compaction;
// also, the block containing the return address may now be dead. Use a
// code_root to take care of the details.
// Allocates memory
cell factor_vm::inline_cache_miss(cell return_address_) {
code_root return_address(return_address_, this);
bool tail_call_site = tail_call_site_p(return_address.value);
#ifdef PIC_DEBUG
FACTOR_PRINT("Inline cache miss at "
<< (tail_call_site ? "tail" : "non-tail")
<< " call site 0x" << std::hex << return_address.value
<< std::dec);
print_callstack();
#endif
data_root<array> cache_entries(ctx->pop(), this);
fixnum index = untag_fixnum(ctx->pop());
data_root<array> methods(ctx->pop(), this);
data_root<word> generic_word(ctx->pop(), this);
data_root<object> object(((cell*)ctx->datastack)[-index], this);
cell pic_size = array_capacity(cache_entries.untagged()) / 2;
update_pic_transitions(pic_size);
cell xt = generic_word->entry_point;
if (pic_size < max_pic_size) {
cell klass = object_class(object.value());
cell method = lookup_method(object.value(), methods.value());
data_root<array> new_cache_entries(
add_inline_cache_entry(cache_entries.value(), klass, method), this);
inline_cache_jit jit(generic_word.value(), this);
jit.emit_inline_cache(index, generic_word.value(), methods.value(),
new_cache_entries.value(), tail_call_site);
code_block* code = jit.to_code_block(CODE_BLOCK_PIC, JIT_FRAME_SIZE);
initialize_code_block(code);
xt = code->entry_point();
}
// Install the new stub.
if (return_address.valid) {
// Since each PIC is only referenced from a single call site,
// if the old call target was a PIC, we can deallocate it immediately,
// instead of leaving dead PICs around until the next GC.
deallocate_inline_cache(return_address.value);
set_call_target(return_address.value, xt);
#ifdef PIC_DEBUG
FACTOR_PRINT("Updated " << (tail_call_site ? "tail" : "non-tail")
<< " call site 0x" << std::hex << return_address.value << std::dec
<< " with 0x" << std::hex << (cell)xt << std::dec);
print_callstack();
#endif
}
return xt;
}
/* Allocates memory */
cell factor_vm::add_inline_cache_entry(cell cache_entries_, cell klass_, cell method_)
{
gc_root<array> cache_entries(cache_entries_,this);
gc_root<object> klass(klass_,this);
gc_root<word> method(method_,this);
cell pic_size = array_capacity(cache_entries.untagged());
gc_root<array> new_cache_entries(reallot_array(cache_entries.untagged(),pic_size + 2),this);
set_array_nth(new_cache_entries.untagged(),pic_size,klass.value());
set_array_nth(new_cache_entries.untagged(),pic_size + 1,method.value());
return new_cache_entries.value();
}
/* The cache_entries parameter is either f (on cold call site) or an array (on cache miss).
Called from assembly with the actual return address */
void *factor_vm::inline_cache_miss(cell return_address)
{
check_code_pointer(return_address);
/* Since each PIC is only referenced from a single call site,
if the old call target was a PIC, we can deallocate it immediately,
instead of leaving dead PICs around until the next GC. */
deallocate_inline_cache(return_address);
gc_root<array> cache_entries(dpop(),this);
fixnum index = untag_fixnum(dpop());
gc_root<array> methods(dpop(),this);
gc_root<word> generic_word(dpop(),this);
gc_root<object> object(((cell *)ds)[-index],this);
void *xt;
cell pic_size = inline_cache_size(cache_entries.value());
update_pic_transitions(pic_size);
if(pic_size >= max_pic_size)
xt = megamorphic_call_stub(generic_word.value());
else
{
cell klass = object_class(object.value());
cell method = lookup_method(object.value(),methods.value());
gc_root<array> new_cache_entries(add_inline_cache_entry(
cache_entries.value(),
klass,
method),this);
xt = compile_inline_cache(index,
generic_word.value(),
methods.value(),
new_cache_entries.value(),
tail_call_site_p(return_address))->xt();
}
/* Install the new stub. */
set_call_target(return_address,xt);
#ifdef PIC_DEBUG
printf("Updated %s call site 0x%lx with 0x%lx\n",
tail_call_site_p(return_address) ? "tail" : "non-tail",
return_address,
(cell)xt);
#endif
return xt;
}
code_block *factor_vm::compile_inline_cache(fixnum index,cell generic_word_,cell methods_,cell cache_entries_,bool tail_call_p)
{
gc_root<word> generic_word(generic_word_,this);
gc_root<array> methods(methods_,this);
gc_root<array> cache_entries(cache_entries_,this);
inline_cache_jit jit(generic_word.value(),this);
jit.compile_inline_cache(index,
generic_word.value(),
methods.value(),
cache_entries.value(),
tail_call_p);
code_block *code = jit.to_code_block();
relocate_code_block(code);
return code;
}
// index: 0 = top of stack, 1 = item underneath, etc
// cache_entries: array of class/method pairs
// Allocates memory
void inline_cache_jit::emit_inline_cache(fixnum index, cell generic_word_,
cell methods_, cell cache_entries_,
bool tail_call_p) {
data_root<word> generic_word(generic_word_, parent);
data_root<array> methods(methods_, parent);
data_root<array> cache_entries(cache_entries_, parent);
cell ic_type = determine_inline_cache_type(cache_entries.untagged());
parent->update_pic_count(ic_type);
// Generate machine code to determine the object's class.
emit_with_literal(parent->special_objects[PIC_LOAD],
tag_fixnum(-index * sizeof(cell)));
// Put the tag of the object, or class of the tuple in a register.
emit(parent->special_objects[ic_type]);
// Generate machine code to check, in turn, if the class is one of the cached
// entries.
for (cell i = 0; i < array_capacity(cache_entries.untagged()); i += 2) {
cell klass = array_nth(cache_entries.untagged(), i);
cell method = array_nth(cache_entries.untagged(), i + 1);
emit_check_and_jump(ic_type, i, klass, method);
}
// If none of the above conditionals tested true, then execution "falls
// through" to here.
// A stack frame is set up, since the inline-cache-miss sub-primitive
// makes a subroutine call to the VM.
emit(parent->special_objects[JIT_PROLOG]);
// The inline-cache-miss sub-primitive call receives enough information to
// reconstruct the PIC with the new entry.
push(generic_word.value());
push(methods.value());
push(tag_fixnum(index));
push(cache_entries.value());
emit_subprimitive(
parent->special_objects[tail_call_p ? PIC_MISS_TAIL_WORD : PIC_MISS_WORD],
true, // tail_call_p
true); // stack_frame_p
}
/* index: 0 = top of stack, 1 = item underneath, etc
cache_entries: array of class/method pairs */
void inline_cache_jit::compile_inline_cache(fixnum index,
cell generic_word_,
cell methods_,
cell cache_entries_,
bool tail_call_p)
{
gc_root<word> generic_word(generic_word_,parent);
gc_root<array> methods(methods_,parent);
gc_root<array> cache_entries(cache_entries_,parent);
cell inline_cache_type = parent->determine_inline_cache_type(cache_entries.untagged());
parent->update_pic_count(inline_cache_type);
/* Generate machine code to determine the object's class. */
emit_class_lookup(index,inline_cache_type);
/* Generate machine code to check, in turn, if the class is one of the cached entries. */
cell i;
for(i = 0; i < array_capacity(cache_entries.untagged()); i += 2)
{
/* Class equal? */
cell klass = array_nth(cache_entries.untagged(),i);
emit_check(klass);
/* Yes? Jump to method */
cell method = array_nth(cache_entries.untagged(),i + 1);
emit_with(parent->userenv[PIC_HIT],method);
}
/* Generate machine code to handle a cache miss, which ultimately results in
this function being called again.
The inline-cache-miss primitive call receives enough information to
reconstruct the PIC. */
push(generic_word.value());
push(methods.value());
push(tag_fixnum(index));
push(cache_entries.value());
word_special(parent->userenv[tail_call_p ? PIC_MISS_TAIL_WORD : PIC_MISS_WORD]);
}