1berry/ruby
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

YJIT: Replace perf_fn! with a simpler macro (#9971)

Browse Source
This commit is contained in:
Takashi Kokubun 2024-02-14 15:57:24 -08:00 committed by GitHub
parent 06ad00adc2
commit beeee54880
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 35 additions and 42 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

77
yjit/src/codegen.rs
View File

@ -319,18 +319,21 @@ macro_rules! jit_perf_symbol_pop {
}; };
} }
/// Macro to push and pop perf symbols around a function definition. /// Macro to push and pop a perf symbol around a function call.
/// This is useful when the function has early returns. macro_rules! perf_call {
macro_rules! perf_fn { // perf_call!("prefix: ", func(...)) uses "prefix: func" as a symbol.
(fn $func_name:ident($jit:ident: $jit_t:ty, $asm:ident: $asm_t:ty, $($arg:ident: $type:ty,)*) -> $ret:ty $block:block) => { ($prefix:expr, $func_name:ident($jit:expr, $asm:expr$(, $arg:expr)*$(,)?) ) => {
fn $func_name($jit: $jit_t, $asm: $asm_t, $($arg: $type),*) -> $ret { {
fn func_body($jit: $jit_t, $asm: $asm_t, $($arg: $type),*) -> $ret $block jit_perf_symbol_push!($jit, $asm, &format!("{}{}", $prefix, stringify!($func_name)), PerfMap::Codegen);
jit_perf_symbol_push!($jit, $asm, stringify!($func_name), PerfMap::Codegen); let ret = $func_name($jit, $asm, $($arg),*);
let ret = func_body($jit, $asm, $($arg),*);
jit_perf_symbol_pop!($jit, $asm, PerfMap::Codegen); jit_perf_symbol_pop!($jit, $asm, PerfMap::Codegen);
ret ret
} }
}; };
// perf_call! { func(...) } uses "func" as a symbol.
{ $func_name:ident($jit:expr, $asm:expr$(, $arg:expr)*$(,)?) } => {
perf_call!("", $func_name($jit, $asm, $($arg),*))
};
} }
use crate::codegen::JCCKinds::*; use crate::codegen::JCCKinds::*;
@ -3352,7 +3355,7 @@ fn gen_opt_neq(
// opt_neq is passed two rb_call_data as arguments: // opt_neq is passed two rb_call_data as arguments:
// first for ==, second for != // first for ==, second for !=
let cd = jit.get_arg(1).as_ptr(); let cd = jit.get_arg(1).as_ptr();
return gen_send_general(jit, asm, ocb, cd, None); perf_call! { gen_send_general(jit, asm, ocb, cd, None) }
} }
fn gen_opt_aref( fn gen_opt_aref(
@ -5994,7 +5997,7 @@ fn gen_push_frame(
asm.mov(cfp_opnd(RUBY_OFFSET_CFP_EP), ep); asm.mov(cfp_opnd(RUBY_OFFSET_CFP_EP), ep);
} }
perf_fn!(fn gen_send_cfunc( fn gen_send_cfunc(
jit: &mut JITState, jit: &mut JITState,
asm: &mut Assembler, asm: &mut Assembler,
ocb: &mut OutlinedCb, ocb: &mut OutlinedCb,
@ -6046,11 +6049,7 @@ perf_fn!(fn gen_send_cfunc(
if kw_arg.is_null() && flags & VM_CALL_OPT_SEND == 0 && flags & VM_CALL_ARGS_SPLAT == 0 && (cfunc_argc == -1 || argc == cfunc_argc) { if kw_arg.is_null() && flags & VM_CALL_OPT_SEND == 0 && flags & VM_CALL_ARGS_SPLAT == 0 && (cfunc_argc == -1 || argc == cfunc_argc) {
let expected_stack_after = asm.ctx.get_stack_size() as i32 - argc; let expected_stack_after = asm.ctx.get_stack_size() as i32 - argc;
if let Some(known_cfunc_codegen) = lookup_cfunc_codegen(unsafe { (*cme).def }) { if let Some(known_cfunc_codegen) = lookup_cfunc_codegen(unsafe { (*cme).def }) {
jit_perf_symbol_push!(jit, asm, "gen_send_cfunc: known_cfunc_codegen", PerfMap::Codegen); if perf_call!("gen_send_cfunc: ", known_cfunc_codegen(jit, asm, ocb, ci, cme, block, argc, recv_known_class)) {
let specialized = known_cfunc_codegen(jit, asm, ocb, ci, cme, block, argc, recv_known_class);
jit_perf_symbol_pop!(jit, asm, PerfMap::Codegen);
if specialized {
assert_eq!(expected_stack_after, asm.ctx.get_stack_size() as i32); assert_eq!(expected_stack_after, asm.ctx.get_stack_size() as i32);
gen_counter_incr(asm, Counter::num_send_cfunc_inline); gen_counter_incr(asm, Counter::num_send_cfunc_inline);
// cfunc codegen generated code. Terminate the block so // cfunc codegen generated code. Terminate the block so
@ -6175,8 +6174,7 @@ perf_fn!(fn gen_send_cfunc(
frame_type |= VM_FRAME_FLAG_CFRAME_KW frame_type |= VM_FRAME_FLAG_CFRAME_KW
} }
jit_perf_symbol_push!(jit, asm, "gen_send_cfunc: gen_push_frame", PerfMap::Codegen); perf_call!("gen_send_cfunc: ", gen_push_frame(jit, asm, ControlFrame {
gen_push_frame(jit, asm, ControlFrame {
frame_type, frame_type,
specval, specval,
cme, cme,
@ -6188,8 +6186,7 @@ perf_fn!(fn gen_send_cfunc(
None // Leave PC uninitialized as cfuncs shouldn't read it None // Leave PC uninitialized as cfuncs shouldn't read it
}, },
iseq: None, iseq: None,
}); }));
jit_perf_symbol_pop!(jit, asm, PerfMap::Codegen);
asm_comment!(asm, "set ec->cfp"); asm_comment!(asm, "set ec->cfp");
let new_cfp = asm.lea(Opnd::mem(64, CFP, -(RUBY_SIZEOF_CONTROL_FRAME as i32))); let new_cfp = asm.lea(Opnd::mem(64, CFP, -(RUBY_SIZEOF_CONTROL_FRAME as i32)));
@ -6281,7 +6278,7 @@ perf_fn!(fn gen_send_cfunc(
// We do this to end the current block after the call // We do this to end the current block after the call
jump_to_next_insn(jit, asm, ocb); jump_to_next_insn(jit, asm, ocb);
Some(EndBlock) Some(EndBlock)
}); }
// Generate RARRAY_LEN. For array_opnd, use Opnd::Reg to reduce memory access, // Generate RARRAY_LEN. For array_opnd, use Opnd::Reg to reduce memory access,
// and use Opnd::Mem to save registers. // and use Opnd::Mem to save registers.
@ -6443,7 +6440,7 @@ fn gen_send_bmethod(
} }
let frame_type = VM_FRAME_MAGIC_BLOCK | VM_FRAME_FLAG_BMETHOD | VM_FRAME_FLAG_LAMBDA; let frame_type = VM_FRAME_MAGIC_BLOCK | VM_FRAME_FLAG_BMETHOD | VM_FRAME_FLAG_LAMBDA;
gen_send_iseq(jit, asm, ocb, iseq, ci, frame_type, Some(capture.ep), cme, block, flags, argc, None) perf_call! { gen_send_iseq(jit, asm, ocb, iseq, ci, frame_type, Some(capture.ep), cme, block, flags, argc, None) }
} }
/// Return the ISEQ's return value if it consists of only putnil/putobject and leave. /// Return the ISEQ's return value if it consists of only putnil/putobject and leave.
@ -6471,7 +6468,7 @@ fn iseq_get_return_value(iseq: IseqPtr) -> Option<VALUE> {
} }
} }
perf_fn!(fn gen_send_iseq( fn gen_send_iseq(
jit: &mut JITState, jit: &mut JITState,
asm: &mut Assembler, asm: &mut Assembler,
ocb: &mut OutlinedCb, ocb: &mut OutlinedCb,
@ -7088,8 +7085,7 @@ perf_fn!(fn gen_send_iseq(
}; };
// Setup the new frame // Setup the new frame
jit_perf_symbol_push!(jit, asm, "gen_send_iseq: gen_push_frame", PerfMap::Codegen); perf_call!("gen_send_iseq: ", gen_push_frame(jit, asm, ControlFrame {
gen_push_frame(jit, asm, ControlFrame {
frame_type, frame_type,
specval, specval,
cme, cme,
@ -7097,8 +7093,7 @@ perf_fn!(fn gen_send_iseq(
sp: callee_sp, sp: callee_sp,
iseq: Some(iseq), iseq: Some(iseq),
pc: None, // We are calling into jitted code, which will set the PC as necessary pc: None, // We are calling into jitted code, which will set the PC as necessary
}); }));
jit_perf_symbol_pop!(jit, asm, PerfMap::Codegen);
// Log the name of the method we're calling to. We intentionally don't do this for inlined ISEQs. // Log the name of the method we're calling to. We intentionally don't do this for inlined ISEQs.
// We also do this after gen_push_frame() to minimize the impact of spill_temps() on asm.ccall(). // We also do this after gen_push_frame() to minimize the impact of spill_temps() on asm.ccall().
@ -7189,7 +7184,7 @@ perf_fn!(fn gen_send_iseq(
); );
Some(EndBlock) Some(EndBlock)
}); }
// Check if we can handle a keyword call // Check if we can handle a keyword call
fn gen_iseq_kw_call_checks( fn gen_iseq_kw_call_checks(
@ -7789,7 +7784,7 @@ fn gen_send_dynamic<F: Fn(&mut Assembler) -> Opnd>(
Some(KeepCompiling) Some(KeepCompiling)
} }
perf_fn!(fn gen_send_general( fn gen_send_general(
jit: &mut JITState, jit: &mut JITState,
asm: &mut Assembler, asm: &mut Assembler,
ocb: &mut OutlinedCb, ocb: &mut OutlinedCb,
@ -7862,8 +7857,7 @@ perf_fn!(fn gen_send_general(
return None; return None;
} }
jit_perf_symbol_push!(jit, asm, "gen_send_general: jit_guard_known_klass", PerfMap::Codegen); perf_call!("gen_send_general: ", jit_guard_known_klass(
jit_guard_known_klass(
jit, jit,
asm, asm,
ocb, ocb,
@ -7873,8 +7867,7 @@ perf_fn!(fn gen_send_general(
comptime_recv, comptime_recv,
SEND_MAX_DEPTH, SEND_MAX_DEPTH,
Counter::guard_send_klass_megamorphic, Counter::guard_send_klass_megamorphic,
); ));
jit_perf_symbol_pop!(jit, asm, PerfMap::Codegen);
// Do method lookup // Do method lookup
let mut cme = unsafe { rb_callable_method_entry(comptime_recv_klass, mid) }; let mut cme = unsafe { rb_callable_method_entry(comptime_recv_klass, mid) };
@ -7925,10 +7918,10 @@ perf_fn!(fn gen_send_general(
VM_METHOD_TYPE_ISEQ => { VM_METHOD_TYPE_ISEQ => {
let iseq = unsafe { get_def_iseq_ptr((*cme).def) }; let iseq = unsafe { get_def_iseq_ptr((*cme).def) };
let frame_type = VM_FRAME_MAGIC_METHOD | VM_ENV_FLAG_LOCAL; let frame_type = VM_FRAME_MAGIC_METHOD | VM_ENV_FLAG_LOCAL;
return gen_send_iseq(jit, asm, ocb, iseq, ci, frame_type, None, cme, block, flags, argc, None); return perf_call! { gen_send_iseq(jit, asm, ocb, iseq, ci, frame_type, None, cme, block, flags, argc, None) };
} }
VM_METHOD_TYPE_CFUNC => { VM_METHOD_TYPE_CFUNC => {
return gen_send_cfunc( return perf_call! { gen_send_cfunc(
jit, jit,
asm, asm,
ocb, ocb,
@ -7938,7 +7931,7 @@ perf_fn!(fn gen_send_general(
Some(comptime_recv_klass), Some(comptime_recv_klass),
flags, flags,
argc, argc,
); ) };
} }
VM_METHOD_TYPE_IVAR => { VM_METHOD_TYPE_IVAR => {
// This is a .send call not supported right now for getters // This is a .send call not supported right now for getters
@ -8242,7 +8235,7 @@ perf_fn!(fn gen_send_general(
} }
} }
} }
}); }
/// Assemble "{class_name}#{method_name}" from a class pointer and a method ID /// Assemble "{class_name}#{method_name}" from a class pointer and a method ID
fn get_method_name(class: Option<VALUE>, mid: u64) -> String { fn get_method_name(class: Option<VALUE>, mid: u64) -> String {
@ -8298,7 +8291,7 @@ fn gen_opt_send_without_block(
) -> Option<CodegenStatus> { ) -> Option<CodegenStatus> {
// Generate specialized code if possible // Generate specialized code if possible
let cd = jit.get_arg(0).as_ptr(); let cd = jit.get_arg(0).as_ptr();
if let Some(status) = gen_send_general(jit, asm, ocb, cd, None) { if let Some(status) = perf_call! { gen_send_general(jit, asm, ocb, cd, None) } {
return Some(status); return Some(status);
} }
@ -8322,7 +8315,7 @@ fn gen_send(
// Generate specialized code if possible // Generate specialized code if possible
let cd = jit.get_arg(0).as_ptr(); let cd = jit.get_arg(0).as_ptr();
let block = jit.get_arg(1).as_optional_ptr().map(|iseq| BlockHandler::BlockISeq(iseq)); let block = jit.get_arg(1).as_optional_ptr().map(|iseq| BlockHandler::BlockISeq(iseq));
if let Some(status) = gen_send_general(jit, asm, ocb, cd, block) { if let Some(status) = perf_call! { gen_send_general(jit, asm, ocb, cd, block) } {
return Some(status); return Some(status);
} }
@ -8435,7 +8428,7 @@ fn gen_invokeblock_specialized(
Counter::guard_invokeblock_iseq_block_changed, Counter::guard_invokeblock_iseq_block_changed,
); );
gen_send_iseq(jit, asm, ocb, comptime_iseq, ci, VM_FRAME_MAGIC_BLOCK, None, 0 as _, None, flags, argc, Some(captured_opnd)) perf_call! { gen_send_iseq(jit, asm, ocb, comptime_iseq, ci, VM_FRAME_MAGIC_BLOCK, None, 0 as _, None, flags, argc, Some(captured_opnd)) }
} else if comptime_handler.0 & 0x3 == 0x3 { // VM_BH_IFUNC_P } else if comptime_handler.0 & 0x3 == 0x3 { // VM_BH_IFUNC_P
// We aren't handling CALLER_SETUP_ARG and CALLER_REMOVE_EMPTY_KW_SPLAT yet. // We aren't handling CALLER_SETUP_ARG and CALLER_REMOVE_EMPTY_KW_SPLAT yet.
if flags & VM_CALL_ARGS_SPLAT != 0 { if flags & VM_CALL_ARGS_SPLAT != 0 {
@ -8648,10 +8641,10 @@ fn gen_invokesuper_specialized(
VM_METHOD_TYPE_ISEQ => { VM_METHOD_TYPE_ISEQ => {
let iseq = unsafe { get_def_iseq_ptr((*cme).def) }; let iseq = unsafe { get_def_iseq_ptr((*cme).def) };
let frame_type = VM_FRAME_MAGIC_METHOD | VM_ENV_FLAG_LOCAL; let frame_type = VM_FRAME_MAGIC_METHOD | VM_ENV_FLAG_LOCAL;
gen_send_iseq(jit, asm, ocb, iseq, ci, frame_type, None, cme, Some(block), ci_flags, argc, None) perf_call! { gen_send_iseq(jit, asm, ocb, iseq, ci, frame_type, None, cme, Some(block), ci_flags, argc, None) }
} }
VM_METHOD_TYPE_CFUNC => { VM_METHOD_TYPE_CFUNC => {
gen_send_cfunc(jit, asm, ocb, ci, cme, Some(block), None, ci_flags, argc) perf_call! { gen_send_cfunc(jit, asm, ocb, ci, cme, Some(block), None, ci_flags, argc) }
} }
_ => unreachable!(), _ => unreachable!(),
} }
@ -8790,7 +8783,7 @@ fn gen_objtostring(
Some(KeepCompiling) Some(KeepCompiling)
} else { } else {
let cd = jit.get_arg(0).as_ptr(); let cd = jit.get_arg(0).as_ptr();
gen_send_general(jit, asm, ocb, cd, None) perf_call! { gen_send_general(jit, asm, ocb, cd, None) }
} }
} }