diff options
Diffstat (limited to 'yjit/src/codegen.rs')
| -rw-r--r-- | yjit/src/codegen.rs | 422 |
1 files changed, 195 insertions, 227 deletions
diff --git a/yjit/src/codegen.rs b/yjit/src/codegen.rs index 7cc4aff473..0fbca85716 100644 --- a/yjit/src/codegen.rs +++ b/yjit/src/codegen.rs @@ -438,7 +438,7 @@ impl<'a> JITState<'a> { fn flush_perf_symbols(&self, cb: &CodeBlock) { assert_eq!(0, self.perf_stack.len()); let path = format!("/tmp/perf-{}.map", std::process::id()); - let mut f = std::fs::File::options().create(true).append(true).open(path).unwrap(); + let mut f = std::io::BufWriter::new(std::fs::File::options().create(true).append(true).open(path).unwrap()); for sym in self.perf_map.borrow().iter() { if let (start, Some(end), name) = sym { // In case the code straddles two pages, part of it belongs to the symbol. @@ -821,11 +821,11 @@ fn gen_stub_exit(ocb: &mut OutlinedCb) -> Option<CodePtr> { /// Generate an exit to return to the interpreter fn gen_exit(exit_pc: *mut VALUE, asm: &mut Assembler) { - #[cfg(all(feature = "disasm", not(test)))] - { + #[cfg(not(test))] + asm_comment!(asm, "exit to interpreter on {}", { let opcode = unsafe { rb_vm_insn_addr2opcode((*exit_pc).as_ptr()) }; - asm_comment!(asm, "exit to interpreter on {}", insn_name(opcode as usize)); - } + insn_name(opcode as usize) + }); if asm.ctx.is_return_landing() { asm.mov(SP, Opnd::mem(64, CFP, RUBY_OFFSET_CFP_SP)); @@ -1094,11 +1094,7 @@ pub fn gen_entry_prologue( let code_ptr = cb.get_write_ptr(); let mut asm = Assembler::new(unsafe { get_iseq_body_local_table_size(iseq) }); - if get_option_ref!(dump_disasm).is_some() { - asm_comment!(asm, "YJIT entry point: {}", iseq_get_location(iseq, 0)); - } else { - asm_comment!(asm, "YJIT entry"); - } + asm_comment!(asm, "YJIT entry point: {}", iseq_get_location(iseq, 0)); asm.frame_setup(); @@ -1212,7 +1208,7 @@ fn gen_check_ints( // Not checking interrupt_mask since it's zero outside finalize_deferred_heap_pages, // signal_exec, or rb_postponed_job_flush. - let interrupt_flag = asm.load(Opnd::mem(32, EC, RUBY_OFFSET_EC_INTERRUPT_FLAG)); + let interrupt_flag = asm.load(Opnd::mem(32, EC, RUBY_OFFSET_EC_INTERRUPT_FLAG as i32)); asm.test(interrupt_flag, interrupt_flag); asm.jnz(Target::side_exit(counter)); @@ -1296,7 +1292,6 @@ pub fn gen_single_block( let mut asm = Assembler::new(jit.num_locals()); asm.ctx = ctx; - #[cfg(feature = "disasm")] if get_option_ref!(dump_disasm).is_some() { let blockid_idx = blockid.idx; let chain_depth = if asm.ctx.get_chain_depth() > 0 { format!("(chain_depth: {})", asm.ctx.get_chain_depth()) } else { "".to_string() }; @@ -2263,7 +2258,8 @@ fn gen_expandarray( let comptime_recv = jit.peek_at_stack(&asm.ctx, 0); - // If the comptime receiver is not an array + // If the comptime receiver is not an array, speculate for when the `rb_check_array_type()` + // conversion returns nil and without side-effects (e.g. arbitrary method calls). if !unsafe { RB_TYPE_P(comptime_recv, RUBY_T_ARRAY) } { // at compile time, ensure to_ary is not defined let target_cme = unsafe { rb_callable_method_entry_or_negative(comptime_recv.class_of(), ID!(to_ary)) }; @@ -2275,13 +2271,19 @@ fn gen_expandarray( return None; } + // Bail when method_missing is defined to avoid generating code to call it. + // Also, for simplicity, bail when BasicObject#method_missing has been removed. + if !assume_method_basic_definition(jit, asm, comptime_recv.class_of(), ID!(method_missing)) { + gen_counter_incr(jit, asm, Counter::expandarray_method_missing); + return None; + } + // invalidate compile block if to_ary is later defined jit.assume_method_lookup_stable(asm, target_cme); jit_guard_known_klass( jit, asm, - comptime_recv.class_of(), array_opnd, array_opnd.into(), comptime_recv, @@ -2311,7 +2313,7 @@ fn gen_expandarray( } // Get the compile-time array length - let comptime_len = unsafe { rb_yjit_array_len(comptime_recv) as u32 }; + let comptime_len = unsafe { rb_jit_array_len(comptime_recv) as u32 }; // Move the array from the stack and check that it's an array. guard_object_is_array( @@ -2516,6 +2518,7 @@ fn gen_setlocal_generic( ep_offset: u32, level: u32, ) -> Option<CodegenStatus> { + // Post condition: The type of of the set local is updated in the Context. let value_type = asm.ctx.get_opnd_type(StackOpnd(0)); // Fallback because of write barrier @@ -2537,6 +2540,11 @@ fn gen_setlocal_generic( ); asm.stack_pop(1); + // Set local type in the context + if level == 0 { + let local_idx = ep_offset_to_local_idx(jit.get_iseq(), ep_offset).as_usize(); + asm.ctx.set_local_type(local_idx, value_type); + } return Some(KeepCompiling); } @@ -2589,6 +2597,7 @@ fn gen_setlocal_generic( ); } + // Set local type in the context if level == 0 { let local_idx = ep_offset_to_local_idx(jit.get_iseq(), ep_offset).as_usize(); asm.ctx.set_local_type(local_idx, value_type); @@ -2753,7 +2762,7 @@ fn gen_checkkeyword( ) -> Option<CodegenStatus> { // When a keyword is unspecified past index 32, a hash will be used // instead. This can only happen in iseqs taking more than 32 keywords. - if unsafe { (*get_iseq_body_param_keyword(jit.iseq)).num >= 32 } { + if unsafe { (*get_iseq_body_param_keyword(jit.iseq)).num >= VM_KW_SPECIFIED_BITS_MAX.try_into().unwrap() } { return None; } @@ -2848,24 +2857,12 @@ fn gen_get_ivar( recv: Opnd, recv_opnd: YARVOpnd, ) -> Option<CodegenStatus> { - let comptime_val_klass = comptime_receiver.class_of(); - // If recv isn't already a register, load it. let recv = match recv { Opnd::InsnOut { .. } => recv, _ => asm.load(recv), }; - // Check if the comptime class uses a custom allocator - let custom_allocator = unsafe { rb_get_alloc_func(comptime_val_klass) }; - let uses_custom_allocator = match custom_allocator { - Some(alloc_fun) => { - let allocate_instance = rb_class_allocate_instance as *const u8; - alloc_fun as *const u8 != allocate_instance - } - None => false, - }; - // Check if the comptime receiver is a T_OBJECT let receiver_t_object = unsafe { RB_TYPE_P(comptime_receiver, RUBY_T_OBJECT) }; // Use a general C call at the last chain to avoid exits on megamorphic shapes @@ -2874,12 +2871,9 @@ fn gen_get_ivar( gen_counter_incr(jit, asm, Counter::num_getivar_megamorphic); } - // If the class uses the default allocator, instances should all be T_OBJECT - // NOTE: This assumes nobody changes the allocator of the class after allocation. - // Eventually, we can encode whether an object is T_OBJECT or not - // inside object shapes. + // NOTE: This assumes T_OBJECT can't ever have the same shape_id as any other type. // too-complex shapes can't use index access, so we use rb_ivar_get for them too. - if !receiver_t_object || uses_custom_allocator || comptime_receiver.shape_too_complex() || megamorphic { + if !comptime_receiver.heap_object_p() || comptime_receiver.shape_too_complex() || megamorphic { // General case. Call rb_ivar_get(). // VALUE rb_ivar_get(VALUE obj, ID id) asm_comment!(asm, "call rb_ivar_get()"); @@ -2904,9 +2898,8 @@ fn gen_get_ivar( let ivar_index = unsafe { let shape_id = comptime_receiver.shape_id_of(); - let shape = rb_shape_lookup(shape_id); - let mut ivar_index: u32 = 0; - if rb_shape_get_iv_index(shape, ivar_name, &mut ivar_index) { + let mut ivar_index: u16 = 0; + if rb_shape_get_iv_index(shape_id, ivar_name, &mut ivar_index) { Some(ivar_index as usize) } else { None @@ -2916,9 +2909,6 @@ fn gen_get_ivar( // Guard heap object (recv_opnd must be used before stack_pop) guard_object_is_heap(asm, recv, recv_opnd, Counter::getivar_not_heap); - // Compile time self is embedded and the ivar index lands within the object - let embed_test_result = unsafe { FL_TEST_RAW(comptime_receiver, VALUE(ROBJECT_EMBED.as_usize())) != VALUE(0) }; - let expected_shape = unsafe { rb_obj_shape_id(comptime_receiver) }; let shape_id_offset = unsafe { rb_shape_id_offset() }; let shape_opnd = Opnd::mem(SHAPE_ID_NUM_BITS as u8, recv, shape_id_offset); @@ -2947,28 +2937,37 @@ fn gen_get_ivar( asm.mov(out_opnd, Qnil.into()); } Some(ivar_index) => { - if embed_test_result { - // See ROBJECT_FIELDS() from include/ruby/internal/core/robject.h - - // Load the variable - let offs = ROBJECT_OFFSET_AS_ARY as i32 + (ivar_index * SIZEOF_VALUE) as i32; - let ivar_opnd = Opnd::mem(64, recv, offs); - - // Push the ivar on the stack - let out_opnd = asm.stack_push(Type::Unknown); - asm.mov(out_opnd, ivar_opnd); + let ivar_opnd = if receiver_t_object { + if comptime_receiver.embedded_p() { + // See ROBJECT_FIELDS() from include/ruby/internal/core/robject.h + + // Load the variable + let offs = ROBJECT_OFFSET_AS_ARY as i32 + (ivar_index * SIZEOF_VALUE) as i32; + Opnd::mem(64, recv, offs) + } else { + // Compile time value is *not* embedded. + + // Get a pointer to the extended table + let tbl_opnd = asm.load(Opnd::mem(64, recv, ROBJECT_OFFSET_AS_HEAP_FIELDS as i32)); + + // Read the ivar from the extended table + Opnd::mem(64, tbl_opnd, (SIZEOF_VALUE * ivar_index) as i32) + } } else { - // Compile time value is *not* embedded. - - // Get a pointer to the extended table - let tbl_opnd = asm.load(Opnd::mem(64, recv, ROBJECT_OFFSET_AS_HEAP_FIELDS as i32)); + asm_comment!(asm, "call rb_ivar_get_at()"); - // Read the ivar from the extended table - let ivar_opnd = Opnd::mem(64, tbl_opnd, (SIZEOF_VALUE * ivar_index) as i32); + if assume_single_ractor_mode(jit, asm) { + asm.ccall(rb_ivar_get_at_no_ractor_check as *const u8, vec![recv, Opnd::UImm((ivar_index as u32).into())]) + } else { + // The function could raise RactorIsolationError. + jit_prepare_non_leaf_call(jit, asm); + asm.ccall(rb_ivar_get_at as *const u8, vec![recv, Opnd::UImm((ivar_index as u32).into()), Opnd::UImm(ivar_name)]) + } + }; - let out_opnd = asm.stack_push(Type::Unknown); - asm.mov(out_opnd, ivar_opnd); - } + // Push the ivar on the stack + let out_opnd = asm.stack_push(Type::Unknown); + asm.mov(out_opnd, ivar_opnd); } } @@ -3074,8 +3073,6 @@ fn gen_set_ivar( recv_opnd: YARVOpnd, ic: Option<*const iseq_inline_iv_cache_entry>, ) -> Option<CodegenStatus> { - let comptime_val_klass = comptime_receiver.class_of(); - // If the comptime receiver is frozen, writing an IV will raise an exception // and we don't want to JIT code to deal with that situation. if comptime_receiver.is_frozen() { @@ -3085,16 +3082,6 @@ fn gen_set_ivar( let stack_type = asm.ctx.get_opnd_type(StackOpnd(0)); - // Check if the comptime class uses a custom allocator - let custom_allocator = unsafe { rb_get_alloc_func(comptime_val_klass) }; - let uses_custom_allocator = match custom_allocator { - Some(alloc_fun) => { - let allocate_instance = rb_class_allocate_instance as *const u8; - alloc_fun as *const u8 != allocate_instance - } - None => false, - }; - // Check if the comptime receiver is a T_OBJECT let receiver_t_object = unsafe { RB_TYPE_P(comptime_receiver, RUBY_T_OBJECT) }; // Use a general C call at the last chain to avoid exits on megamorphic shapes @@ -3105,11 +3092,10 @@ fn gen_set_ivar( // Get the iv index let shape_too_complex = comptime_receiver.shape_too_complex(); - let ivar_index = if !shape_too_complex { + let ivar_index = if !comptime_receiver.special_const_p() && !shape_too_complex { let shape_id = comptime_receiver.shape_id_of(); - let shape = unsafe { rb_shape_lookup(shape_id) }; - let mut ivar_index: u32 = 0; - if unsafe { rb_shape_get_iv_index(shape, ivar_name, &mut ivar_index) } { + let mut ivar_index: u16 = 0; + if unsafe { rb_shape_get_iv_index(shape_id, ivar_name, &mut ivar_index) } { Some(ivar_index as usize) } else { None @@ -3121,27 +3107,29 @@ fn gen_set_ivar( // The current shape doesn't contain this iv, we need to transition to another shape. let mut new_shape_too_complex = false; let new_shape = if !shape_too_complex && receiver_t_object && ivar_index.is_none() { - let current_shape = comptime_receiver.shape_of(); - let next_shape_id = unsafe { rb_shape_transition_add_ivar_no_warnings(comptime_receiver, ivar_name) }; - let next_shape = unsafe { rb_shape_lookup(next_shape_id) }; + let current_shape_id = comptime_receiver.shape_id_of(); + // We don't need to check about imemo_fields here because we're definitely looking at a T_OBJECT. + let klass = unsafe { rb_obj_class(comptime_receiver) }; + let next_shape_id = unsafe { rb_shape_transition_add_ivar_no_warnings(klass, current_shape_id, ivar_name) }; // If the VM ran out of shapes, or this class generated too many leaf, // it may be de-optimized into OBJ_TOO_COMPLEX_SHAPE (hash-table). - new_shape_too_complex = unsafe { rb_shape_too_complex_p(next_shape) }; + new_shape_too_complex = unsafe { rb_jit_shape_too_complex_p(next_shape_id) }; if new_shape_too_complex { Some((next_shape_id, None, 0_usize)) } else { - let current_capacity = unsafe { (*current_shape).capacity }; + let current_capacity = unsafe { rb_yjit_shape_capacity(current_shape_id) }; + let next_capacity = unsafe { rb_yjit_shape_capacity(next_shape_id) }; // If the new shape has a different capacity, or is TOO_COMPLEX, we'll have to // reallocate it. - let needs_extension = unsafe { (*current_shape).capacity != (*next_shape).capacity }; + let needs_extension = next_capacity != current_capacity; // We can write to the object, but we need to transition the shape - let ivar_index = unsafe { (*current_shape).next_field_index } as usize; + let ivar_index = unsafe { rb_yjit_shape_index(next_shape_id) } as usize; let needs_extension = if needs_extension { - Some((current_capacity, unsafe { (*next_shape).capacity })) + Some((current_capacity, next_capacity)) } else { None }; @@ -3151,10 +3139,9 @@ fn gen_set_ivar( None }; - // If the receiver isn't a T_OBJECT, or uses a custom allocator, - // then just write out the IV write as a function call. + // If the receiver isn't a T_OBJECT, then just write out the IV write as a function call. // too-complex shapes can't use index access, so we use rb_ivar_get for them too. - if !receiver_t_object || uses_custom_allocator || shape_too_complex || new_shape_too_complex || megamorphic { + if !receiver_t_object || shape_too_complex || new_shape_too_complex || megamorphic { // The function could raise FrozenError. // Note that this modifies REG_SP, which is why we do it first jit_prepare_non_leaf_call(jit, asm); @@ -3178,7 +3165,7 @@ fn gen_set_ivar( asm.ccall( rb_vm_setinstancevariable as *const u8, vec![ - Opnd::const_ptr(jit.iseq as *const u8), + VALUE(jit.iseq as usize).into(), Opnd::mem(64, CFP, RUBY_OFFSET_CFP_SELF), ivar_name.into(), val_opnd, @@ -3371,7 +3358,7 @@ fn gen_definedivar( // Specialize base on compile time values let comptime_receiver = jit.peek_at_self(); - if comptime_receiver.shape_too_complex() || asm.ctx.get_chain_depth() >= GET_IVAR_MAX_DEPTH { + if comptime_receiver.special_const_p() || comptime_receiver.shape_too_complex() || asm.ctx.get_chain_depth() >= GET_IVAR_MAX_DEPTH { // Fall back to calling rb_ivar_defined // Save the PC and SP because the callee may allocate @@ -3397,9 +3384,8 @@ fn gen_definedivar( let shape_id = comptime_receiver.shape_id_of(); let ivar_exists = unsafe { - let shape = rb_shape_lookup(shape_id); - let mut ivar_index: u32 = 0; - rb_shape_get_iv_index(shape, ivar_name, &mut ivar_index) + let mut ivar_index: u16 = 0; + rb_shape_get_iv_index(shape_id, ivar_name, &mut ivar_index) }; // Guard heap object (recv_opnd must be used before stack_pop) @@ -3697,7 +3683,6 @@ fn gen_equality_specialized( jit_guard_known_klass( jit, asm, - unsafe { rb_cString }, a_opnd, a_opnd.into(), comptime_a, @@ -3723,7 +3708,6 @@ fn gen_equality_specialized( jit_guard_known_klass( jit, asm, - unsafe { rb_cString }, b_opnd, b_opnd.into(), comptime_b, @@ -3820,7 +3804,6 @@ fn gen_opt_aref( jit_guard_known_klass( jit, asm, - unsafe { rb_cArray }, recv_opnd, recv_opnd.into(), comptime_recv, @@ -3860,7 +3843,6 @@ fn gen_opt_aref( jit_guard_known_klass( jit, asm, - unsafe { rb_cHash }, recv_opnd, recv_opnd.into(), comptime_recv, @@ -3891,40 +3873,6 @@ fn gen_opt_aref( } } -fn gen_opt_aset_with( - jit: &mut JITState, - asm: &mut Assembler, -) -> Option<CodegenStatus> { - // We might allocate or raise - jit_prepare_non_leaf_call(jit, asm); - - let key_opnd = Opnd::Value(jit.get_arg(0)); - let recv_opnd = asm.stack_opnd(1); - let value_opnd = asm.stack_opnd(0); - - extern "C" { - fn rb_vm_opt_aset_with(recv: VALUE, key: VALUE, value: VALUE) -> VALUE; - } - - let val_opnd = asm.ccall( - rb_vm_opt_aset_with as *const u8, - vec![ - recv_opnd, - key_opnd, - value_opnd, - ], - ); - asm.stack_pop(2); // Keep it on stack during GC - - asm.cmp(val_opnd, Qundef.into()); - asm.je(Target::side_exit(Counter::opt_aset_with_qundef)); - - let top = asm.stack_push(Type::Unknown); - asm.mov(top, val_opnd); - - return Some(KeepCompiling); -} - fn gen_opt_aset( jit: &mut JITState, asm: &mut Assembler, @@ -3947,7 +3895,6 @@ fn gen_opt_aset( jit_guard_known_klass( jit, asm, - unsafe { rb_cArray }, recv, recv.into(), comptime_recv, @@ -3959,7 +3906,6 @@ fn gen_opt_aset( jit_guard_known_klass( jit, asm, - unsafe { rb_cInteger }, key, key.into(), comptime_key, @@ -3992,7 +3938,6 @@ fn gen_opt_aset( jit_guard_known_klass( jit, asm, - unsafe { rb_cHash }, recv, recv.into(), comptime_recv, @@ -4020,38 +3965,6 @@ fn gen_opt_aset( } } -fn gen_opt_aref_with( - jit: &mut JITState, - asm: &mut Assembler, -) -> Option<CodegenStatus>{ - // We might allocate or raise - jit_prepare_non_leaf_call(jit, asm); - - let key_opnd = Opnd::Value(jit.get_arg(0)); - let recv_opnd = asm.stack_opnd(0); - - extern "C" { - fn rb_vm_opt_aref_with(recv: VALUE, key: VALUE) -> VALUE; - } - - let val_opnd = asm.ccall( - rb_vm_opt_aref_with as *const u8, - vec![ - recv_opnd, - key_opnd - ], - ); - asm.stack_pop(1); // Keep it on stack during GC - - asm.cmp(val_opnd, Qundef.into()); - asm.je(Target::side_exit(Counter::opt_aref_with_qundef)); - - let top = asm.stack_push(Type::Unknown); - asm.mov(top, val_opnd); - - return Some(KeepCompiling); -} - fn gen_opt_and( jit: &mut JITState, asm: &mut Assembler, @@ -4318,11 +4231,11 @@ fn gen_opt_ary_freeze( return None; } - let str = jit.get_arg(0); + let ary = jit.get_arg(0); // Push the return value onto the stack let stack_ret = asm.stack_push(Type::CArray); - asm.mov(stack_ret, str.into()); + asm.mov(stack_ret, ary.into()); Some(KeepCompiling) } @@ -4335,11 +4248,11 @@ fn gen_opt_hash_freeze( return None; } - let str = jit.get_arg(0); + let hash = jit.get_arg(0); // Push the return value onto the stack let stack_ret = asm.stack_push(Type::CHash); - asm.mov(stack_ret, str.into()); + asm.mov(stack_ret, hash.into()); Some(KeepCompiling) } @@ -4944,7 +4857,6 @@ fn gen_opt_new( perf_call!("opt_new: ", jit_guard_known_klass( jit, asm, - comptime_recv_klass, recv, recv.into(), comptime_recv, @@ -5015,13 +4927,13 @@ fn gen_jump( fn jit_guard_known_klass( jit: &mut JITState, asm: &mut Assembler, - known_klass: VALUE, obj_opnd: Opnd, insn_opnd: YARVOpnd, sample_instance: VALUE, max_chain_depth: u8, counter: Counter, ) { + let known_klass = sample_instance.class_of(); let val_type = asm.ctx.get_opnd_type(insn_opnd); if val_type.known_class() == Some(known_klass) { @@ -5127,7 +5039,7 @@ fn jit_guard_known_klass( assert_eq!(sample_instance.class_of(), rb_cString, "context says class is exactly ::String") }; } else { - assert!(!val_type.is_imm()); + assert!(!val_type.is_imm(), "{insn_opnd:?} should be a heap object, but was {val_type:?} for {sample_instance:?}"); // Check that the receiver is a heap object // Note: if we get here, the class doesn't have immediate instances. @@ -5771,7 +5683,6 @@ fn jit_rb_float_plus( jit_guard_known_klass( jit, asm, - comptime_obj.class_of(), obj, obj.into(), comptime_obj, @@ -5813,7 +5724,6 @@ fn jit_rb_float_minus( jit_guard_known_klass( jit, asm, - comptime_obj.class_of(), obj, obj.into(), comptime_obj, @@ -5855,7 +5765,6 @@ fn jit_rb_float_mul( jit_guard_known_klass( jit, asm, - comptime_obj.class_of(), obj, obj.into(), comptime_obj, @@ -5897,7 +5806,6 @@ fn jit_rb_float_div( jit_guard_known_klass( jit, asm, - comptime_obj.class_of(), obj, obj.into(), comptime_obj, @@ -6161,7 +6069,6 @@ fn jit_rb_str_getbyte( jit_guard_known_klass( jit, asm, - comptime_idx.class_of(), idx, idx.into(), comptime_idx, @@ -6278,16 +6185,19 @@ fn jit_rb_str_dup( jit_prepare_call_with_gc(jit, asm); - // Check !FL_ANY_RAW(str, FL_EXIVAR), which is part of BARE_STRING_P. - let recv_opnd = asm.stack_pop(1); + let recv_opnd = asm.stack_opnd(0); let recv_opnd = asm.load(recv_opnd); - let flags_opnd = Opnd::mem(64, recv_opnd, RUBY_OFFSET_RBASIC_FLAGS); - asm.test(flags_opnd, Opnd::Imm(RUBY_FL_EXIVAR as i64)); + + let shape_id_offset = unsafe { rb_shape_id_offset() }; + let shape_opnd = Opnd::mem(64, recv_opnd, shape_id_offset); + asm.test(shape_opnd, Opnd::UImm(SHAPE_ID_HAS_IVAR_MASK as u64)); asm.jnz(Target::side_exit(Counter::send_str_dup_exivar)); // Call rb_str_dup - let stack_ret = asm.stack_push(Type::CString); let ret_opnd = asm.ccall(rb_str_dup as *const u8, vec![recv_opnd]); + + asm.stack_pop(1); + let stack_ret = asm.stack_push(Type::CString); asm.mov(stack_ret, ret_opnd); true @@ -6343,7 +6253,7 @@ fn jit_rb_str_concat_codepoint( guard_object_is_fixnum(jit, asm, codepoint, StackOpnd(0)); - asm.ccall(rb_yjit_str_concat_codepoint as *const u8, vec![recv, codepoint]); + asm.ccall(rb_jit_str_concat_codepoint as *const u8, vec![recv, codepoint]); // The receiver is the return value, so we only need to pop the codepoint argument off the stack. // We can reuse the receiver slot in the stack as the return value. @@ -6649,6 +6559,7 @@ fn jit_rb_f_block_given_p( true } +/// Codegen for `block_given?` and `defined?(yield)` fn gen_block_given( jit: &mut JITState, asm: &mut Assembler, @@ -6658,16 +6569,24 @@ fn gen_block_given( ) { asm_comment!(asm, "block_given?"); - // Same as rb_vm_frame_block_handler - let ep_opnd = gen_get_lep(jit, asm); - let block_handler = asm.load( - Opnd::mem(64, ep_opnd, SIZEOF_VALUE_I32 * VM_ENV_DATA_INDEX_SPECVAL) - ); + // `yield` goes to the block handler stowed in the "local" iseq which is + // the current iseq or a parent. Only the "method" iseq type can be passed a + // block handler. (e.g. `yield` in the top level script is a syntax error.) + let local_iseq = unsafe { rb_get_iseq_body_local_iseq(jit.iseq) }; + if unsafe { rb_get_iseq_body_type(local_iseq) } == ISEQ_TYPE_METHOD { + // Same as rb_vm_frame_block_handler + let ep_opnd = gen_get_lep(jit, asm); + let block_handler = asm.load( + Opnd::mem(64, ep_opnd, SIZEOF_VALUE_I32 * VM_ENV_DATA_INDEX_SPECVAL) + ); - // Return `block_handler != VM_BLOCK_HANDLER_NONE` - asm.cmp(block_handler, VM_BLOCK_HANDLER_NONE.into()); - let block_given = asm.csel_ne(true_opnd, false_opnd); - asm.mov(out_opnd, block_given); + // Return `block_handler != VM_BLOCK_HANDLER_NONE` + asm.cmp(block_handler, VM_BLOCK_HANDLER_NONE.into()); + let block_given = asm.csel_ne(true_opnd, false_opnd); + asm.mov(out_opnd, block_given); + } else { + asm.mov(out_opnd, false_opnd); + } } // Codegen for rb_class_superclass() @@ -6740,7 +6659,7 @@ fn jit_thread_s_current( asm.stack_pop(1); // ec->thread_ptr - let ec_thread_opnd = asm.load(Opnd::mem(64, EC, RUBY_OFFSET_EC_THREAD_PTR)); + let ec_thread_opnd = asm.load(Opnd::mem(64, EC, RUBY_OFFSET_EC_THREAD_PTR as i32)); // thread->self let thread_self = Opnd::mem(64, ec_thread_opnd, RUBY_OFFSET_THREAD_SELF); @@ -7205,7 +7124,7 @@ fn gen_send_cfunc( asm_comment!(asm, "set ec->cfp"); let new_cfp = asm.lea(Opnd::mem(64, CFP, -(RUBY_SIZEOF_CONTROL_FRAME as i32))); - asm.store(Opnd::mem(64, EC, RUBY_OFFSET_EC_CFP), new_cfp); + asm.store(Opnd::mem(64, EC, RUBY_OFFSET_EC_CFP as i32), new_cfp); if !kw_arg.is_null() { // Build a hash from all kwargs passed @@ -7301,7 +7220,7 @@ fn gen_send_cfunc( // Pop the stack frame (ec->cfp++) // Instead of recalculating, we can reuse the previous CFP, which is stored in a callee-saved // register - let ec_cfp_opnd = Opnd::mem(64, EC, RUBY_OFFSET_EC_CFP); + let ec_cfp_opnd = Opnd::mem(64, EC, RUBY_OFFSET_EC_CFP as i32); asm.store(ec_cfp_opnd, CFP); // cfunc calls may corrupt types @@ -7467,7 +7386,7 @@ fn gen_send_bmethod( ) -> Option<CodegenStatus> { let procv = unsafe { rb_get_def_bmethod_proc((*cme).def) }; - let proc = unsafe { rb_yjit_get_proc_ptr(procv) }; + let proc = unsafe { rb_jit_get_proc_ptr(procv) }; let proc_block = unsafe { &(*proc).block }; if proc_block.type_ != block_type_iseq { @@ -7477,11 +7396,12 @@ fn gen_send_bmethod( let capture = unsafe { proc_block.as_.captured.as_ref() }; let iseq = unsafe { *capture.code.iseq.as_ref() }; - // Optimize for single ractor mode and avoid runtime check for - // "defined with an un-shareable Proc in a different Ractor" - if !assume_single_ractor_mode(jit, asm) { - gen_counter_incr(jit, asm, Counter::send_bmethod_ractor); - return None; + if !procv.shareable_p() { + let ractor_serial = unsafe { rb_yjit_cme_ractor_serial(cme) }; + asm_comment!(asm, "guard current ractor == {}", ractor_serial); + let current_ractor_serial = asm.load(Opnd::mem(64, EC, RUBY_OFFSET_EC_RACTOR_ID as i32)); + asm.cmp(current_ractor_serial, ractor_serial.into()); + asm.jne(Target::side_exit(Counter::send_bmethod_ractor)); } // Passing a block to a block needs logic different from passing @@ -7547,6 +7467,12 @@ fn iseq_get_return_value(iseq: IseqPtr, captured_opnd: Option<Opnd>, block: Opti let ep_offset = unsafe { *rb_iseq_pc_at_idx(iseq, 1) }.as_u32(); let local_idx = ep_offset_to_local_idx(iseq, ep_offset); + // Only inline getlocal on a parameter. DCE in the IESQ builder can + // make a two-instruction ISEQ that does not return a parameter. + if local_idx >= unsafe { get_iseq_body_param_size(iseq) } { + return None; + } + if unsafe { rb_simple_iseq_p(iseq) } { return Some(IseqReturn::LocalVariable(local_idx)); } else if unsafe { rb_iseq_only_kwparam_p(iseq) } { @@ -7703,7 +7629,7 @@ fn gen_send_iseq( gen_counter_incr(jit, asm, Counter::send_iseq_splat_not_array); return None; } else { - unsafe { rb_yjit_array_len(array) as u32} + unsafe { rb_jit_array_len(array) as u32} }; // Arity check accounting for size of the splat. When callee has rest parameters, we insert @@ -7794,7 +7720,7 @@ fn gen_send_iseq( gen_counter_incr(jit, asm, Counter::num_send_iseq); // Shortcut for special `Primitive.attr! :leaf` builtins - let builtin_attrs = unsafe { rb_yjit_iseq_builtin_attrs(iseq) }; + let builtin_attrs = unsafe { rb_jit_iseq_builtin_attrs(iseq) }; let builtin_func_raw = unsafe { rb_yjit_builtin_function(iseq) }; let builtin_func = if builtin_func_raw.is_null() { None } else { Some(builtin_func_raw) }; let opt_send_call = flags & VM_CALL_OPT_SEND != 0; // .send call is not currently supported for builtins @@ -7971,6 +7897,11 @@ fn gen_send_iseq( gen_counter_incr(jit, asm, Counter::send_iseq_clobbering_block_arg); return None; } + if iseq_has_rest || has_kwrest { + // The proc would be stored above the current stack top, where GC can't see it + gen_counter_incr(jit, asm, Counter::send_iseq_block_arg_gc_unsafe); + return None; + } let proc = asm.stack_pop(1); // Pop first, as argc doesn't account for the block arg let callee_specval = asm.ctx.sp_opnd(callee_specval); asm.store(callee_specval, proc); @@ -8383,7 +8314,7 @@ fn gen_send_iseq( // We also do this after spill_regs() to avoid doubly spilling the same thing on asm.ccall(). if get_option!(gen_stats) { // Protect caller-saved registers in case they're used for arguments - asm.cpush_all(); + let mapping = asm.cpush_all(); // Assemble the ISEQ name string let name_str = get_iseq_name(iseq); @@ -8393,7 +8324,7 @@ fn gen_send_iseq( // Increment the counter for this cfunc asm.ccall(incr_iseq_counter as *const u8, vec![iseq_idx.into()]); - asm.cpop_all(); + asm.cpop_all(mapping); } // The callee might change locals through Kernel#binding and other means. @@ -8428,7 +8359,7 @@ fn gen_send_iseq( asm_comment!(asm, "switch to new CFP"); let new_cfp = asm.sub(CFP, RUBY_SIZEOF_CONTROL_FRAME.into()); asm.mov(CFP, new_cfp); - asm.store(Opnd::mem(64, EC, RUBY_OFFSET_EC_CFP), CFP); + asm.store(Opnd::mem(64, EC, RUBY_OFFSET_EC_CFP as i32), CFP); // Directly jump to the entry point of the callee gen_direct_jump( @@ -9016,6 +8947,12 @@ fn gen_struct_aset( return None; } + // If the comptime receiver is frozen, writing a struct member will raise an exception + // and we don't want to JIT code to deal with that situation. + if comptime_recv.is_frozen() { + return None; + } + if c_method_tracing_currently_enabled(jit) { // Struct accesses need fire c_call and c_return events, which we can't support // See :attr-tracing: @@ -9036,6 +8973,17 @@ fn gen_struct_aset( assert!(unsafe { RB_TYPE_P(comptime_recv, RUBY_T_STRUCT) }); assert!((off as i64) < unsafe { RSTRUCT_LEN(comptime_recv) }); + // Even if the comptime recv was not frozen, future recv may be. So we need to emit a guard + // that the recv is not frozen. + // We know all structs are heap objects, so we can check the flag directly. + let recv = asm.stack_opnd(1); + let recv = asm.load(recv); + let flags = asm.load(Opnd::mem(VALUE_BITS, recv, RUBY_OFFSET_RBASIC_FLAGS)); + asm.test(flags, (RUBY_FL_FREEZE as u64).into()); + asm.jnz(Target::side_exit(Counter::opt_aset_frozen)); + + // Not frozen, so we can proceed. + asm_comment!(asm, "struct aset"); let val = asm.stack_pop(1); @@ -9147,7 +9095,6 @@ fn gen_send_general( let recv_opnd: YARVOpnd = recv.into(); // Log the name of the method we're calling to - #[cfg(feature = "disasm")] asm_comment!(asm, "call to {}", get_method_name(Some(comptime_recv_klass), mid)); // Gather some statistics about sends @@ -9167,7 +9114,6 @@ fn gen_send_general( perf_call!("gen_send_general: ", jit_guard_known_klass( jit, asm, - comptime_recv_klass, recv, recv_opnd, comptime_recv, @@ -9435,13 +9381,6 @@ fn gen_send_general( return None; } - // Optimize for single ractor mode and avoid runtime check for - // "defined with an un-shareable Proc in a different Ractor" - if !assume_single_ractor_mode(jit, asm) { - gen_counter_incr(jit, asm, Counter::send_call_multi_ractor); - return None; - } - // If this is a .send call we need to adjust the stack if flags & VM_CALL_OPT_SEND != 0 { handle_opt_send_shift_stack(asm, argc); @@ -9646,7 +9585,24 @@ fn gen_sendforward( jit: &mut JITState, asm: &mut Assembler, ) -> Option<CodegenStatus> { - return gen_send(jit, asm); + // Generate specialized code if possible + let cd = jit.get_arg(0).as_ptr(); + let block = jit.get_arg(1).as_optional_ptr().map(|iseq| BlockHandler::BlockISeq(iseq)); + if let Some(status) = perf_call! { gen_send_general(jit, asm, cd, block) } { + return Some(status); + } + + // Otherwise, fallback to dynamic dispatch using the interpreter's implementation of sendforward + let blockiseq = jit.get_arg(1).as_iseq(); + gen_send_dynamic(jit, asm, cd, unsafe { rb_yjit_sendish_sp_pops((*cd).ci) }, |asm| { + extern "C" { + fn rb_vm_sendforward(ec: EcPtr, cfp: CfpPtr, cd: VALUE, blockiseq: IseqPtr) -> VALUE; + } + asm.ccall( + rb_vm_sendforward as *const u8, + vec![EC, CFP, (cd as usize).into(), VALUE(blockiseq as usize).into()], + ) + }) } fn gen_invokeblock( @@ -9720,7 +9676,7 @@ fn gen_invokeblock_specialized( // If the current ISEQ is annotated to be inlined but it's not being inlined here, // generate a dynamic dispatch to avoid making this yield megamorphic. - if unsafe { rb_yjit_iseq_builtin_attrs(jit.iseq) } & BUILTIN_ATTR_INLINE_BLOCK != 0 && !asm.ctx.inline() { + if unsafe { rb_jit_iseq_builtin_attrs(jit.iseq) } & BUILTIN_ATTR_INLINE_BLOCK != 0 && !asm.ctx.inline() { gen_counter_incr(jit, asm, Counter::invokeblock_iseq_not_inlined); return None; } @@ -9811,7 +9767,7 @@ fn gen_invokesuper( return Some(status); } - // Otherwise, fallback to dynamic dispatch using the interpreter's implementation of send + // Otherwise, fallback to dynamic dispatch using the interpreter's implementation of invokesuper let blockiseq = jit.get_arg(1).as_iseq(); gen_send_dynamic(jit, asm, cd, unsafe { rb_yjit_sendish_sp_pops((*cd).ci) }, |asm| { extern "C" { @@ -9828,7 +9784,23 @@ fn gen_invokesuperforward( jit: &mut JITState, asm: &mut Assembler, ) -> Option<CodegenStatus> { - return gen_invokesuper(jit, asm); + // Generate specialized code if possible + let cd = jit.get_arg(0).as_ptr(); + if let Some(status) = gen_invokesuper_specialized(jit, asm, cd) { + return Some(status); + } + + // Otherwise, fallback to dynamic dispatch using the interpreter's implementation of invokesuperforward + let blockiseq = jit.get_arg(1).as_iseq(); + gen_send_dynamic(jit, asm, cd, unsafe { rb_yjit_sendish_sp_pops((*cd).ci) }, |asm| { + extern "C" { + fn rb_vm_invokesuperforward(ec: EcPtr, cfp: CfpPtr, cd: VALUE, blockiseq: IseqPtr) -> VALUE; + } + asm.ccall( + rb_vm_invokesuperforward as *const u8, + vec![EC, CFP, (cd as usize).into(), VALUE(blockiseq as usize).into()], + ) + }) } fn gen_invokesuper_specialized( @@ -9982,7 +9954,7 @@ fn gen_leave( asm_comment!(asm, "pop stack frame"); let incr_cfp = asm.add(CFP, RUBY_SIZEOF_CONTROL_FRAME.into()); asm.mov(CFP, incr_cfp); - asm.mov(Opnd::mem(64, EC, RUBY_OFFSET_EC_CFP), CFP); + asm.mov(Opnd::mem(64, EC, RUBY_OFFSET_EC_CFP as i32), CFP); // Load the return value let retval_opnd = asm.stack_pop(1); @@ -10081,7 +10053,6 @@ fn gen_objtostring( jit_guard_known_klass( jit, asm, - comptime_recv.class_of(), recv, recv.into(), comptime_recv, @@ -10095,7 +10066,6 @@ fn gen_objtostring( jit_guard_known_klass( jit, asm, - comptime_recv.class_of(), recv, recv.into(), comptime_recv, @@ -10279,7 +10249,7 @@ fn gen_getclassvariable( let val_opnd = asm.ccall( rb_vm_getclassvariable as *const u8, vec![ - Opnd::mem(64, CFP, RUBY_OFFSET_CFP_ISEQ), + VALUE(jit.iseq as usize).into(), CFP, Opnd::UImm(jit.get_arg(0).as_u64()), Opnd::UImm(jit.get_arg(1).as_u64()), @@ -10303,7 +10273,7 @@ fn gen_setclassvariable( asm.ccall( rb_vm_setclassvariable as *const u8, vec![ - Opnd::mem(64, CFP, RUBY_OFFSET_CFP_ISEQ), + VALUE(jit.iseq as usize).into(), CFP, Opnd::UImm(jit.get_arg(0).as_u64()), val, @@ -10786,8 +10756,6 @@ fn get_gen_fn(opcode: VALUE) -> Option<InsnGenFn> { YARVINSN_opt_neq => Some(gen_opt_neq), YARVINSN_opt_aref => Some(gen_opt_aref), YARVINSN_opt_aset => Some(gen_opt_aset), - YARVINSN_opt_aref_with => Some(gen_opt_aref_with), - YARVINSN_opt_aset_with => Some(gen_opt_aset_with), YARVINSN_opt_mult => Some(gen_opt_mult), YARVINSN_opt_div => Some(gen_opt_div), YARVINSN_opt_ltlt => Some(gen_opt_ltlt), @@ -11019,7 +10987,7 @@ impl CodegenGlobals { #[cfg(not(test))] let (mut cb, mut ocb) = { - let virt_block: *mut u8 = unsafe { rb_yjit_reserve_addr_space(exec_mem_size as u32) }; + let virt_block: *mut u8 = unsafe { rb_jit_reserve_addr_space(exec_mem_size as u32) }; // Memory protection syscalls need page-aligned addresses, so check it here. Assuming // `virt_block` is page-aligned, `second_half` should be page-aligned as long as the @@ -11028,7 +10996,7 @@ impl CodegenGlobals { // // Basically, we don't support x86-64 2MiB and 1GiB pages. ARMv8 can do up to 64KiB // (2¹⁶ bytes) pages, which should be fine. 4KiB pages seem to be the most popular though. - let page_size = unsafe { rb_yjit_get_page_size() }; + let page_size = unsafe { rb_jit_get_page_size() }; assert_eq!( virt_block as usize % page_size.as_usize(), 0, "Start of virtual address block should be page-aligned", @@ -11044,7 +11012,7 @@ impl CodegenGlobals { exec_mem_size, get_option!(mem_size), ); - let mem_block = Rc::new(RefCell::new(mem_block)); + let mem_block = Rc::new(mem_block); let freed_pages = Rc::new(None); |