1
0
Fork 0
mirror of https://github.com/ruby/ruby.git synced 2022-11-09 12:17:21 -05:00
ruby--ruby/ujit_iface.c

412 lines
12 KiB
C
Raw Normal View History

#include <assert.h>
#include "insns.inc"
#include "internal.h"
#include "vm_core.h"
#include "vm_sync.h"
#include "vm_callinfo.h"
#include "builtin.h"
#include "internal/compile.h"
#include "internal/class.h"
#include "insns_info.inc"
#include "ujit.h"
#include "ujit_iface.h"
#include "ujit_codegen.h"
#include "ujit_core.h"
#include "ujit_hooks.inc"
Expose methods for inspecting Micro JIT code blocks This commit adds a module `UJIT`. The module allows you to insert the initial Micro JIT instruction in to an arbitrary iseq like this: ```ruby def foo(x) if x < 1 "less than one" else "something else" end end iseq = RubyVM::InstructionSequence.of(method(:foo)) UJIT.insert(iseq) # Add initial jump ``` After the initial jump is added, we can make Micro JIT do some work: ```ruby 100.times { foo(0) } ``` The `UJIT` module also exposes a method for finding all compiled blocks for a given iseq, like this: ```ruby blocks = UJIT.blocks_for(iseq) ``` We can sort the blocks by address and use the Crabstone gem (which is a wrapper around `capstone`) to disassemble the generated code. Here is the full code example: ```ruby def foo(x) if x < 1 "less than one" else "something else" end end iseq = RubyVM::InstructionSequence.of(method(:foo)) UJIT.insert(iseq) # Add initial jump 100.times { foo(0) } blocks = UJIT.blocks_for(iseq) # brew install capstone # gem install crabstone require "crabstone" cs = Crabstone::Disassembler.new(Crabstone::ARCH_X86, Crabstone::MODE_64) puts iseq.disasm blocks.sort_by(&:address).reverse.each do |block| puts "== ISEQ RANGE: #{block.iseq_start_index} -> #{block.iseq_end_index} ".ljust(80, "=") cs.disasm(block.code, 0).each do |i| printf( "\t0x%<address>x:\t%<instruction>s\t%<details>s\n", address: i.address, instruction: i.mnemonic, details: i.op_str ) end end ``` Here is the output: ``` $ ./ruby test.rb == disasm: #<ISeq:foo@test.rb:1 (1,0)-(7,3)> (catch: FALSE) local table (size: 1, argc: 1 [opts: 0, rest: -1, post: 0, block: -1, kw: -1@-1, kwrest: -1]) [ 1] x@0<Arg> 0000 getlocal_WC_0 x@0 ( 2)[LiCa] 0002 putobject_INT2FIX_1_ 0003 opt_lt <calldata!mid:<, argc:1, ARGS_SIMPLE> 0005 branchunless 10 0007 putstring "less than one" ( 3)[Li] 0009 leave ( 7)[Re] 0010 putstring "something else" ( 5)[Li] 0012 leave ( 7)[Re] == ISEQ RANGE: 7 -> 7 ========================================================== 0x0: movabs rax, 0x7fcd014cd518 0xa: mov qword ptr [rdi], rax 0xd: mov r8, rax 0x10: mov r9, rax 0x13: mov r11, r12 0x16: jmp qword ptr [rax] == ISEQ RANGE: 0 -> 7 ========================================================== 0x0: mov rax, qword ptr [rdi + 0x20] 0x4: mov rax, qword ptr [rax - 0x18] 0x8: mov qword ptr [rdx], rax 0xb: mov qword ptr [rdx + 8], 3 0x13: movabs rax, 0x7fcd0180ac00 0x1d: test byte ptr [rax + 0x3e6], 1 0x24: jne 0x3ffe0da 0x2a: test byte ptr [rdx], 1 0x2d: je 0x3ffe0da 0x33: test byte ptr [rdx + 8], 1 0x37: je 0x3ffe0da 0x3d: mov rax, qword ptr [rdx] 0x40: cmp rax, qword ptr [rdx + 8] 0x44: movabs rax, 0 0x4e: movabs rcx, 0x14 0x58: cmovl rax, rcx 0x5c: mov qword ptr [rdx], rax 0x5f: test qword ptr [rdx], -9 0x66: je 0x3ffe111 0x6c: jmp 0xffffffffffffffa3 ```
2021-01-20 13:50:13 -05:00
VALUE cUjitBlock;
extern st_table * version_tbl;
extern codeblock_t *cb;
static const rb_data_type_t ujit_block_type = {
"UJIT/Block",
{0, 0, 0, },
0, 0, RUBY_TYPED_FREE_IMMEDIATELY
};
bool rb_ujit_enabled;
// Hash table of encoded instructions
extern st_table *rb_encoded_insn_data;
2020-12-08 17:19:28 -05:00
// Write the uJIT entry point pre-call bytes
void
cb_write_pre_call_bytes(codeblock_t* cb)
{
for (size_t i = 0; i < sizeof(ujit_with_ec_pre_call_bytes); ++i)
cb_write_byte(cb, ujit_with_ec_pre_call_bytes[i]);
}
// Write the uJIT exit post-call bytes
void
cb_write_post_call_bytes(codeblock_t* cb)
{
for (size_t i = 0; i < sizeof(ujit_with_ec_post_call_bytes); ++i)
cb_write_byte(cb, ujit_with_ec_post_call_bytes[i]);
}
// Keep track of mapping from instructions to generated code
// See comment for rb_encoded_insn_data in iseq.c
void
map_addr2insn(void *code_ptr, int insn)
{
const void * const *table = rb_vm_get_insns_address_table();
const void * const translated_address = table[insn];
st_data_t encoded_insn_data;
if (st_lookup(rb_encoded_insn_data, (st_data_t)translated_address, &encoded_insn_data)) {
st_insert(rb_encoded_insn_data, (st_data_t)code_ptr, encoded_insn_data);
}
else {
rb_bug("ujit: failed to find info for original instruction while dealing with addr2insn");
}
}
int
opcode_at_pc(const rb_iseq_t *iseq, const VALUE *pc)
{
const VALUE at_pc = *pc;
if (FL_TEST_RAW((VALUE)iseq, ISEQ_TRANSLATED)) {
return rb_vm_insn_addr2opcode((const void *)at_pc);
}
else {
return (int)at_pc;
}
}
// Verify that calling with cd on receiver goes to callee
void
check_cfunc_dispatch(VALUE receiver, struct rb_call_data *cd, void *callee, rb_callable_method_entry_t *compile_time_cme)
{
if (METHOD_ENTRY_INVALIDATED(compile_time_cme)) {
rb_bug("ujit: output code uses invalidated cme %p", (void *)compile_time_cme);
}
bool callee_correct = false;
const rb_callable_method_entry_t *cme = rb_callable_method_entry(CLASS_OF(receiver), vm_ci_mid(cd->ci));
if (cme->def->type == VM_METHOD_TYPE_CFUNC) {
const rb_method_cfunc_t *cfunc = UNALIGNED_MEMBER_PTR(cme->def, body.cfunc);
if ((void *)cfunc->func == callee) {
callee_correct = true;
}
}
if (!callee_correct) {
rb_bug("ujit: output code calls wrong method cd->cc->klass: %p", (void *)cd->cc->klass);
}
}
MJIT_FUNC_EXPORTED VALUE rb_hash_has_key(VALUE hash, VALUE key);
bool
cfunc_needs_frame(const rb_method_cfunc_t *cfunc)
{
void* fptr = (void*)cfunc->func;
// Leaf C functions do not need a stack frame
// or a stack overflow check
return !(
// Hash#key?
fptr == (void*)rb_hash_has_key
);
}
// GC root for interacting with the GC
struct ujit_root_struct {};
// Map cme_or_cc => [[iseq, offset]]. An entry in the map means compiled code at iseq[offset]
// is only valid when cme_or_cc is valid
static st_table *method_lookup_dependency;
struct compiled_region_array {
int32_t size;
int32_t capa;
block_t* data[];
};
// Add an element to a region array, or allocate a new region array.
static struct compiled_region_array *
add_compiled_region(struct compiled_region_array *array, block_t* block)
{
if (!array) {
// Allocate a brand new array with space for one
array = malloc(sizeof(*array) + sizeof(block_t*));
if (!array) {
return NULL;
}
array->size = 0;
array->capa = 1;
}
if (array->size == INT32_MAX) {
return NULL;
}
// Check if the region is already present
for (int32_t i = 0; i < array->size; i++) {
if (array->data[i] == block) {
return array;
}
}
if (array->size + 1 > array->capa) {
// Double the array's capacity.
int64_t double_capa = ((int64_t)array->capa) * 2;
int32_t new_capa = (int32_t)double_capa;
if (new_capa != double_capa) {
return NULL;
}
array = realloc(array, sizeof(*array) + new_capa * sizeof(block_t*));
if (array == NULL) {
return NULL;
}
array->capa = new_capa;
}
int32_t size = array->size;
array->data[size] = block;
array->size++;
return array;
}
static int
add_lookup_dependency_i(st_data_t *key, st_data_t *value, st_data_t data, int existing)
{
block_t *block = (block_t *)data;
struct compiled_region_array *regions = NULL;
if (existing) {
regions = (struct compiled_region_array *)*value;
}
regions = add_compiled_region(regions, block);
if (!regions) {
rb_bug("ujit: failed to add method lookup dependency"); // TODO: we could bail out of compiling instead
}
*value = (st_data_t)regions;
return ST_CONTINUE;
}
// Remember that the currently compiling region is only valid while cme and cc are valid
void
assume_method_lookup_stable(const struct rb_callcache *cc, const rb_callable_method_entry_t *cme, block_t* block)
{
assert (block != NULL);
st_update(method_lookup_dependency, (st_data_t)cme, add_lookup_dependency_i, (st_data_t)block);
st_update(method_lookup_dependency, (st_data_t)cc, add_lookup_dependency_i, (st_data_t)block);
// FIXME: This is a leak! When either the cme or the cc become invalid, the other also needs to go
}
static int
ujit_root_mark_i(st_data_t k, st_data_t v, st_data_t ignore)
{
// FIXME: This leaks everything that end up in the dependency table!
// One way to deal with this is with weak references...
rb_gc_mark((VALUE)k);
struct compiled_region_array *regions = (void *)v;
for (int32_t i = 0; i < regions->size; i++) {
rb_gc_mark((VALUE)regions->data[i]->blockid.iseq);
}
return ST_CONTINUE;
}
// GC callback during mark phase
static void
ujit_root_mark(void *ptr)
{
if (method_lookup_dependency) {
st_foreach(method_lookup_dependency, ujit_root_mark_i, 0);
}
}
static void
ujit_root_free(void *ptr)
{
// Do nothing. The root lives as long as the process.
}
static size_t
ujit_root_memsize(const void *ptr)
{
// Count off-gc-heap allocation size of the dependency table
return st_memsize(method_lookup_dependency); // TODO: more accurate accounting
}
// Custom type for interacting with the GC
// TODO: compaction support
// TODO: make this write barrier protected
static const rb_data_type_t ujit_root_type = {
"ujit_root",
{ujit_root_mark, ujit_root_free, ujit_root_memsize, },
0, 0, RUBY_TYPED_FREE_IMMEDIATELY
};
// Callback when cme or cc become invalid
void
rb_ujit_method_lookup_change(VALUE cme_or_cc)
{
if (!method_lookup_dependency)
return;
RUBY_ASSERT(IMEMO_TYPE_P(cme_or_cc, imemo_ment) || IMEMO_TYPE_P(cme_or_cc, imemo_callcache));
st_data_t image;
if (st_lookup(method_lookup_dependency, (st_data_t)cme_or_cc, &image)) {
struct compiled_region_array *array = (void *)image;
// Invalidate all regions that depend on the cme or cc
for (int32_t i = 0; i < array->size; i++) {
block_t* block = array->data[i];
/*
struct compiled_region *region = &array->data[i];
const struct rb_iseq_constant_body *body = region->iseq->body;
RUBY_ASSERT((unsigned int)region->start_idx < body->iseq_size);
// Restore region address to interpreter address in bytecode sequence
if (body->iseq_encoded[region->start_idx] == (VALUE)region->code) {
const void *const *code_threading_table = rb_vm_get_insns_address_table();
int opcode = rb_vm_insn_addr2insn(region->code);
body->iseq_encoded[region->start_idx] = (VALUE)code_threading_table[opcode];
if (UJIT_DUMP_MODE > 0) {
fprintf(stderr, "cc_or_cme=%p now out of date. Restored idx=%u in iseq=%p\n", (void *)cme_or_cc, (unsigned)region->start_idx, (void *)region->iseq);
}
}
*/
invalidate(block);
}
array->size = 0;
}
}
void
rb_ujit_compile_iseq(const rb_iseq_t *iseq)
{
#if OPT_DIRECT_THREADED_CODE || OPT_CALL_THREADED_CODE
RB_VM_LOCK_ENTER();
VALUE *encoded = (VALUE *)iseq->body->iseq_encoded;
2020-12-10 00:06:10 -05:00
// Compile a block version starting at the first instruction
uint8_t* code_ptr = gen_entry_point(iseq, 0);
if (code_ptr)
{
// Map the code address to the corresponding opcode
int first_opcode = opcode_at_pc(iseq, &encoded[0]);
map_addr2insn(code_ptr, first_opcode);
encoded[0] = (VALUE)code_ptr;
}
2020-12-10 00:06:10 -05:00
RB_VM_LOCK_LEAVE();
#endif
}
Expose methods for inspecting Micro JIT code blocks This commit adds a module `UJIT`. The module allows you to insert the initial Micro JIT instruction in to an arbitrary iseq like this: ```ruby def foo(x) if x < 1 "less than one" else "something else" end end iseq = RubyVM::InstructionSequence.of(method(:foo)) UJIT.insert(iseq) # Add initial jump ``` After the initial jump is added, we can make Micro JIT do some work: ```ruby 100.times { foo(0) } ``` The `UJIT` module also exposes a method for finding all compiled blocks for a given iseq, like this: ```ruby blocks = UJIT.blocks_for(iseq) ``` We can sort the blocks by address and use the Crabstone gem (which is a wrapper around `capstone`) to disassemble the generated code. Here is the full code example: ```ruby def foo(x) if x < 1 "less than one" else "something else" end end iseq = RubyVM::InstructionSequence.of(method(:foo)) UJIT.insert(iseq) # Add initial jump 100.times { foo(0) } blocks = UJIT.blocks_for(iseq) # brew install capstone # gem install crabstone require "crabstone" cs = Crabstone::Disassembler.new(Crabstone::ARCH_X86, Crabstone::MODE_64) puts iseq.disasm blocks.sort_by(&:address).reverse.each do |block| puts "== ISEQ RANGE: #{block.iseq_start_index} -> #{block.iseq_end_index} ".ljust(80, "=") cs.disasm(block.code, 0).each do |i| printf( "\t0x%<address>x:\t%<instruction>s\t%<details>s\n", address: i.address, instruction: i.mnemonic, details: i.op_str ) end end ``` Here is the output: ``` $ ./ruby test.rb == disasm: #<ISeq:foo@test.rb:1 (1,0)-(7,3)> (catch: FALSE) local table (size: 1, argc: 1 [opts: 0, rest: -1, post: 0, block: -1, kw: -1@-1, kwrest: -1]) [ 1] x@0<Arg> 0000 getlocal_WC_0 x@0 ( 2)[LiCa] 0002 putobject_INT2FIX_1_ 0003 opt_lt <calldata!mid:<, argc:1, ARGS_SIMPLE> 0005 branchunless 10 0007 putstring "less than one" ( 3)[Li] 0009 leave ( 7)[Re] 0010 putstring "something else" ( 5)[Li] 0012 leave ( 7)[Re] == ISEQ RANGE: 7 -> 7 ========================================================== 0x0: movabs rax, 0x7fcd014cd518 0xa: mov qword ptr [rdi], rax 0xd: mov r8, rax 0x10: mov r9, rax 0x13: mov r11, r12 0x16: jmp qword ptr [rax] == ISEQ RANGE: 0 -> 7 ========================================================== 0x0: mov rax, qword ptr [rdi + 0x20] 0x4: mov rax, qword ptr [rax - 0x18] 0x8: mov qword ptr [rdx], rax 0xb: mov qword ptr [rdx + 8], 3 0x13: movabs rax, 0x7fcd0180ac00 0x1d: test byte ptr [rax + 0x3e6], 1 0x24: jne 0x3ffe0da 0x2a: test byte ptr [rdx], 1 0x2d: je 0x3ffe0da 0x33: test byte ptr [rdx + 8], 1 0x37: je 0x3ffe0da 0x3d: mov rax, qword ptr [rdx] 0x40: cmp rax, qword ptr [rdx + 8] 0x44: movabs rax, 0 0x4e: movabs rcx, 0x14 0x58: cmovl rax, rcx 0x5c: mov qword ptr [rdx], rax 0x5f: test qword ptr [rdx], -9 0x66: je 0x3ffe111 0x6c: jmp 0xffffffffffffffa3 ```
2021-01-20 13:50:13 -05:00
struct ujit_block_itr {
const rb_iseq_t *iseq;
VALUE list;
};
static int
iseqw_ujit_collect_blocks(st_data_t key, st_data_t value, st_data_t argp)
{
block_t * block = (block_t *)value;
struct ujit_block_itr * itr = (struct ujit_block_itr *)argp;
if (block->blockid.iseq == itr->iseq) {
VALUE rb_block = TypedData_Wrap_Struct(cUjitBlock, &ujit_block_type, block);
rb_ary_push(itr->list, rb_block);
}
return ST_CONTINUE;
}
/* Get a list of the UJIT blocks associated with `rb_iseq` */
static VALUE
ujit_blocks_for(VALUE mod, VALUE rb_iseq)
{
const rb_iseq_t *iseq = rb_iseqw_to_iseq(rb_iseq);
st_table * vt = (st_table *)version_tbl;
struct ujit_block_itr itr;
itr.iseq = iseq;
itr.list = rb_ary_new();
rb_st_foreach(vt, iseqw_ujit_collect_blocks, (st_data_t)&itr);
return itr.list;
}
static VALUE
ujit_insert(VALUE mod, VALUE iseq)
{
rb_ujit_compile_iseq(rb_iseqw_to_iseq(iseq));
return iseq;
}
/* Get the address of the UJIT::Block */
static VALUE
block_address(VALUE self)
{
block_t * block;
TypedData_Get_Struct(self, block_t, &ujit_block_type, block);
return LONG2NUM((intptr_t)block);
}
/* Get the machine code for UJIT::Block as a binary string */
static VALUE
block_code(VALUE self)
{
block_t * block;
TypedData_Get_Struct(self, block_t, &ujit_block_type, block);
return rb_str_new(cb->mem_block + block->start_pos, block->end_pos - block->start_pos);
}
/* Get the start index in the Instruction Sequence that corresponds to this
* UJIT::Block */
static VALUE
iseq_start_index(VALUE self)
{
block_t * block;
TypedData_Get_Struct(self, block_t, &ujit_block_type, block);
return INT2NUM(block->blockid.idx);
}
/* Get the end index in the Instruction Sequence that corresponds to this
* UJIT::Block */
static VALUE
iseq_end_index(VALUE self)
{
block_t * block;
TypedData_Get_Struct(self, block_t, &ujit_block_type, block);
return INT2NUM(block->end_idx);
}
void
rb_ujit_init(void)
{
if (!ujit_scrape_successful || !PLATFORM_SUPPORTED_P)
{
return;
}
rb_ujit_enabled = true;
2020-12-10 16:59:13 -05:00
ujit_init_core();
ujit_init_codegen();
Expose methods for inspecting Micro JIT code blocks This commit adds a module `UJIT`. The module allows you to insert the initial Micro JIT instruction in to an arbitrary iseq like this: ```ruby def foo(x) if x < 1 "less than one" else "something else" end end iseq = RubyVM::InstructionSequence.of(method(:foo)) UJIT.insert(iseq) # Add initial jump ``` After the initial jump is added, we can make Micro JIT do some work: ```ruby 100.times { foo(0) } ``` The `UJIT` module also exposes a method for finding all compiled blocks for a given iseq, like this: ```ruby blocks = UJIT.blocks_for(iseq) ``` We can sort the blocks by address and use the Crabstone gem (which is a wrapper around `capstone`) to disassemble the generated code. Here is the full code example: ```ruby def foo(x) if x < 1 "less than one" else "something else" end end iseq = RubyVM::InstructionSequence.of(method(:foo)) UJIT.insert(iseq) # Add initial jump 100.times { foo(0) } blocks = UJIT.blocks_for(iseq) # brew install capstone # gem install crabstone require "crabstone" cs = Crabstone::Disassembler.new(Crabstone::ARCH_X86, Crabstone::MODE_64) puts iseq.disasm blocks.sort_by(&:address).reverse.each do |block| puts "== ISEQ RANGE: #{block.iseq_start_index} -> #{block.iseq_end_index} ".ljust(80, "=") cs.disasm(block.code, 0).each do |i| printf( "\t0x%<address>x:\t%<instruction>s\t%<details>s\n", address: i.address, instruction: i.mnemonic, details: i.op_str ) end end ``` Here is the output: ``` $ ./ruby test.rb == disasm: #<ISeq:foo@test.rb:1 (1,0)-(7,3)> (catch: FALSE) local table (size: 1, argc: 1 [opts: 0, rest: -1, post: 0, block: -1, kw: -1@-1, kwrest: -1]) [ 1] x@0<Arg> 0000 getlocal_WC_0 x@0 ( 2)[LiCa] 0002 putobject_INT2FIX_1_ 0003 opt_lt <calldata!mid:<, argc:1, ARGS_SIMPLE> 0005 branchunless 10 0007 putstring "less than one" ( 3)[Li] 0009 leave ( 7)[Re] 0010 putstring "something else" ( 5)[Li] 0012 leave ( 7)[Re] == ISEQ RANGE: 7 -> 7 ========================================================== 0x0: movabs rax, 0x7fcd014cd518 0xa: mov qword ptr [rdi], rax 0xd: mov r8, rax 0x10: mov r9, rax 0x13: mov r11, r12 0x16: jmp qword ptr [rax] == ISEQ RANGE: 0 -> 7 ========================================================== 0x0: mov rax, qword ptr [rdi + 0x20] 0x4: mov rax, qword ptr [rax - 0x18] 0x8: mov qword ptr [rdx], rax 0xb: mov qword ptr [rdx + 8], 3 0x13: movabs rax, 0x7fcd0180ac00 0x1d: test byte ptr [rax + 0x3e6], 1 0x24: jne 0x3ffe0da 0x2a: test byte ptr [rdx], 1 0x2d: je 0x3ffe0da 0x33: test byte ptr [rdx + 8], 1 0x37: je 0x3ffe0da 0x3d: mov rax, qword ptr [rdx] 0x40: cmp rax, qword ptr [rdx + 8] 0x44: movabs rax, 0 0x4e: movabs rcx, 0x14 0x58: cmovl rax, rcx 0x5c: mov qword ptr [rdx], rax 0x5f: test qword ptr [rdx], -9 0x66: je 0x3ffe111 0x6c: jmp 0xffffffffffffffa3 ```
2021-01-20 13:50:13 -05:00
VALUE mUjit = rb_define_module("UJIT");
rb_define_module_function(mUjit, "install_entry", ujit_insert, 1);
rb_define_module_function(mUjit, "blocks_for", ujit_blocks_for, 1);
cUjitBlock = rb_define_class_under(mUjit, "Block", rb_cObject);
rb_define_method(cUjitBlock, "address", block_address, 0);
rb_define_method(cUjitBlock, "code", block_code, 0);
rb_define_method(cUjitBlock, "iseq_start_index", iseq_start_index, 0);
rb_define_method(cUjitBlock, "iseq_end_index", iseq_end_index, 0);
// Initialize the GC hooks
method_lookup_dependency = st_init_numtable();
struct ujit_root_struct *root;
VALUE ujit_root = TypedData_Make_Struct(0, struct ujit_root_struct, &ujit_root_type, root);
rb_gc_register_mark_object(ujit_root);
}