YJIT: Optimize local variables when EP == BP (#10487)

vm.c

@@ -1007,6 +1007,11 @@ vm_make_env_each(const rb_execution_context_t * const ec, rb_control_frame_t *co
     }
 #endif
 
+    // Invalidate JIT code that assumes cfp->ep == vm_base_ptr(cfp).
+    if (env->iseq) {
+        rb_yjit_invalidate_ep_is_bp(env->iseq);
+    }
+
     return (VALUE)env;
 }
 

yjit.c

@@ -629,6 +629,12 @@ rb_get_iseq_body_stack_max(const rb_iseq_t *iseq)
     return iseq->body->stack_max;
 }
 
+enum rb_iseq_type
+rb_get_iseq_body_type(const rb_iseq_t *iseq)
+{
+    return iseq->body->type;
+}
+
 bool
 rb_get_iseq_flags_has_lead(const rb_iseq_t *iseq)
 {

yjit.h

@@ -48,6 +48,7 @@ void rb_yjit_tracing_invalidate_all(void);
 void rb_yjit_show_usage(int help, int highlight, unsigned int width, int columns);
 void rb_yjit_lazy_push_frame(const VALUE *pc);
 void rb_yjit_invalidate_no_singleton_class(VALUE klass);
+void rb_yjit_invalidate_ep_is_bp(const rb_iseq_t *iseq);
 
 #else
 // !USE_YJIT
@@ -71,6 +72,7 @@ static inline void rb_yjit_constant_ic_update(const rb_iseq_t *const iseq, IC ic
 static inline void rb_yjit_tracing_invalidate_all(void) {}
 static inline void rb_yjit_lazy_push_frame(const VALUE *pc) {}
 static inline void rb_yjit_invalidate_no_singleton_class(VALUE klass) {}
+static inline void rb_yjit_invalidate_ep_is_bp(const rb_iseq_t *iseq) {}
 
 #endif // #if USE_YJIT
 

yjit/bindgen/src/main.rs

@@ -298,6 +298,7 @@ fn main() {
         .allowlist_type("ruby_tag_type")
         .allowlist_type("ruby_vm_throw_flags")
         .allowlist_type("vm_check_match_type")
+        .allowlist_type("rb_iseq_type")
 
         // From yjit.c
         .allowlist_function("rb_iseq_(get|set)_yjit_payload")
@@ -415,6 +416,7 @@ fn main() {
         .allowlist_function("rb_get_iseq_body_parent_iseq")
         .allowlist_function("rb_get_iseq_body_iseq_encoded")
         .allowlist_function("rb_get_iseq_body_stack_max")
+        .allowlist_function("rb_get_iseq_body_type")
         .allowlist_function("rb_get_iseq_flags_has_lead")
         .allowlist_function("rb_get_iseq_flags_has_opt")
         .allowlist_function("rb_get_iseq_flags_has_kw")

yjit/src/codegen.rs

@@ -46,7 +46,7 @@ type InsnGenFn = fn(
 /// Represents a [core::Block] while we build it.
 pub struct JITState {
     /// Instruction sequence for the compiling block
-    iseq: IseqPtr,
+    pub iseq: IseqPtr,
 
     /// The iseq index of the first instruction in the block
     starting_insn_idx: IseqIdx,
@@ -101,6 +101,9 @@ pub struct JITState {
     /// A list of classes that are not supposed to have a singleton class.
     pub no_singleton_class_assumptions: Vec<VALUE>,
 
+    /// When true, the block is valid only when base pointer is equal to environment pointer.
+    pub no_ep_escape: bool,
+
     /// When true, the block is valid only when there is a total of one ractor running
     pub block_assumes_single_ractor: bool,
 
@@ -130,6 +133,7 @@ impl JITState {
             bop_assumptions: vec![],
             stable_constant_names_assumption: None,
             no_singleton_class_assumptions: vec![],
+            no_ep_escape: false,
             block_assumes_single_ractor: false,
             perf_map: Rc::default(),
             perf_stack: vec![],
@@ -171,6 +175,23 @@ impl JITState {
         unsafe { *(self.pc.offset(arg_idx + 1)) }
     }
 
+    /// Return true if the current ISEQ could escape an environment.
+    ///
+    /// As of vm_push_frame(), EP is always equal to BP. However, after pushing
+    /// a frame, some ISEQ setups call vm_bind_update_env(), which redirects EP.
+    /// Also, some method calls escape the environment to the heap.
+    fn escapes_ep(&self) -> bool {
+        match unsafe { get_iseq_body_type(self.iseq) } {
+            // <main> frame is always associated to TOPLEVEL_BINDING.
+            ISEQ_TYPE_MAIN |
+            // Kernel#eval uses a heap EP when a Binding argument is not nil.
+            ISEQ_TYPE_EVAL => true,
+            // If this ISEQ has previously escaped EP, give up the optimization.
+            _ if iseq_escapes_ep(self.iseq) => true,
+            _ => false,
+        }
+    }
+
     // Get the index of the next instruction
     fn next_insn_idx(&self) -> u16 {
         self.insn_idx + insn_len(self.get_opcode()) as u16
@@ -250,6 +271,19 @@ impl JITState {
         true
     }
 
+    /// Assume that base pointer is equal to environment pointer in the current ISEQ.
+    /// Return true if it's safe to assume so.
+    fn assume_no_ep_escape(&mut self, asm: &mut Assembler, ocb: &mut OutlinedCb) -> bool {
+        if jit_ensure_block_entry_exit(self, asm, ocb).is_none() {
+            return false; // out of space, give up
+        }
+        if self.escapes_ep() {
+            return false; // EP has been escaped in this ISEQ. disable the optimization to avoid an invalidation loop.
+        }
+        self.no_ep_escape = true;
+        true
+    }
+
     fn get_cfp(&self) -> *mut rb_control_frame_struct {
         unsafe { get_ec_cfp(self.ec) }
     }
@@ -2203,16 +2237,22 @@ fn gen_get_lep(jit: &JITState, asm: &mut Assembler) -> Opnd {
 fn gen_getlocal_generic(
     jit: &mut JITState,
     asm: &mut Assembler,
+    ocb: &mut OutlinedCb,
     ep_offset: u32,
     level: u32,
 ) -> Option<CodegenStatus> {
-    // Load environment pointer EP (level 0) from CFP
-    let ep_opnd = gen_get_ep(asm, level);
+    let local_opnd = if level == 0 && jit.assume_no_ep_escape(asm, ocb) {
+        // Load the local using SP register
+        asm.ctx.ep_opnd(-(ep_offset as i32))
+    } else {
+        // Load environment pointer EP (level 0) from CFP
+        let ep_opnd = gen_get_ep(asm, level);
 
-    // Load the local from the block
-    // val = *(vm_get_ep(GET_EP(), level) - idx);
-    let offs = -(SIZEOF_VALUE_I32 * ep_offset as i32);
-    let local_opnd = Opnd::mem(64, ep_opnd, offs);
+        // Load the local from the block
+        // val = *(vm_get_ep(GET_EP(), level) - idx);
+        let offs = -(SIZEOF_VALUE_I32 * ep_offset as i32);
+        Opnd::mem(64, ep_opnd, offs)
+    };
 
     // Write the local at SP
     let stack_top = if level == 0 {
@@ -2230,29 +2270,29 @@ fn gen_getlocal_generic(
 fn gen_getlocal(
     jit: &mut JITState,
     asm: &mut Assembler,
-    _ocb: &mut OutlinedCb,
+    ocb: &mut OutlinedCb,
 ) -> Option<CodegenStatus> {
     let idx = jit.get_arg(0).as_u32();
     let level = jit.get_arg(1).as_u32();
-    gen_getlocal_generic(jit, asm, idx, level)
+    gen_getlocal_generic(jit, asm, ocb, idx, level)
 }
 
 fn gen_getlocal_wc0(
     jit: &mut JITState,
     asm: &mut Assembler,
-    _ocb: &mut OutlinedCb,
+    ocb: &mut OutlinedCb,
 ) -> Option<CodegenStatus> {
     let idx = jit.get_arg(0).as_u32();
-    gen_getlocal_generic(jit, asm, idx, 0)
+    gen_getlocal_generic(jit, asm, ocb, idx, 0)
 }
 
 fn gen_getlocal_wc1(
     jit: &mut JITState,
     asm: &mut Assembler,
-    _ocb: &mut OutlinedCb,
+    ocb: &mut OutlinedCb,
 ) -> Option<CodegenStatus> {
     let idx = jit.get_arg(0).as_u32();
-    gen_getlocal_generic(jit, asm, idx, 1)
+    gen_getlocal_generic(jit, asm, ocb, idx, 1)
 }
 
 fn gen_setlocal_generic(
@@ -2264,11 +2304,11 @@ fn gen_setlocal_generic(
 ) -> Option<CodegenStatus> {
     let value_type = asm.ctx.get_opnd_type(StackOpnd(0));
 
-    // Load environment pointer EP at level
-    let ep_opnd = gen_get_ep(asm, level);
-
     // Fallback because of write barrier
     if asm.ctx.get_chain_depth() > 0 {
+        // Load environment pointer EP at level
+        let ep_opnd = gen_get_ep(asm, level);
+
         // This function should not yield to the GC.
         // void rb_vm_env_write(const VALUE *ep, int index, VALUE v)
         let index = -(ep_offset as i64);
@@ -2286,16 +2326,27 @@ fn gen_setlocal_generic(
         return Some(KeepCompiling);
     }
 
-    // Write barriers may be required when VM_ENV_FLAG_WB_REQUIRED is set, however write barriers
-    // only affect heap objects being written. If we know an immediate value is being written we
-    // can skip this check.
-    if !value_type.is_imm() {
-        // flags & VM_ENV_FLAG_WB_REQUIRED
+    let (flags_opnd, local_opnd) = if level == 0 && jit.assume_no_ep_escape(asm, ocb) {
+        // Load flags and the local using SP register
+        let local_opnd = asm.ctx.ep_opnd(-(ep_offset as i32));
+        let flags_opnd = asm.ctx.ep_opnd(VM_ENV_DATA_INDEX_FLAGS as i32);
+        (flags_opnd, local_opnd)
+    } else {
+        // Load flags and the local for the level
+        let ep_opnd = gen_get_ep(asm, level);
         let flags_opnd = Opnd::mem(
             64,
             ep_opnd,
             SIZEOF_VALUE_I32 * VM_ENV_DATA_INDEX_FLAGS as i32,
         );
+        (flags_opnd, Opnd::mem(64, ep_opnd, -SIZEOF_VALUE_I32 * ep_offset as i32))
+    };
+
+    // Write barriers may be required when VM_ENV_FLAG_WB_REQUIRED is set, however write barriers
+    // only affect heap objects being written. If we know an immediate value is being written we
+    // can skip this check.
+    if !value_type.is_imm() {
+        // flags & VM_ENV_FLAG_WB_REQUIRED
         asm.test(flags_opnd, VM_ENV_FLAG_WB_REQUIRED.into());
 
         // if (flags & VM_ENV_FLAG_WB_REQUIRED) != 0
@@ -2319,8 +2370,7 @@ fn gen_setlocal_generic(
     let stack_top = asm.stack_pop(1);
 
     // Write the value at the environment pointer
-    let offs = -(SIZEOF_VALUE_I32 * ep_offset as i32);
-    asm.mov(Opnd::mem(64, ep_opnd, offs), stack_top);
+    asm.mov(local_opnd, stack_top);
 
     Some(KeepCompiling)
 }

yjit/src/core.rs

@@ -1657,6 +1657,9 @@ impl JITState {
         for klass in self.no_singleton_class_assumptions {
             track_no_singleton_class_assumption(blockref, klass);
         }
+        if self.no_ep_escape {
+            track_no_ep_escape_assumption(blockref, self.iseq);
+        }
 
         blockref
     }
@@ -1798,6 +1801,13 @@ impl Context {
         return Opnd::mem(64, SP, offset);
     }
 
+    /// Get an operand for the adjusted environment pointer address using SP register.
+    /// This is valid only when a Binding object hasn't been created for the frame.
+    pub fn ep_opnd(&self, offset: i32) -> Opnd {
+        let ep_offset = self.get_stack_size() as i32 + 1;
+        self.sp_opnd(-ep_offset + offset)
+    }
+
     /// Stop using a register for a given stack temp.
     /// This allows us to reuse the register for a value that we know is dead
     /// and will no longer be used (e.g. popped stack temp).
@@ -3124,6 +3134,12 @@ pub fn defer_compilation(
     // Likely a stub due to the increased chain depth
     let target0_address = branch.set_target(0, blockid, &next_ctx, ocb);
 
+    // Pad the block if it has the potential to be invalidated. This must be
+    // done before gen_fn() in case the jump is overwritten by a fallthrough.
+    if jit.block_entry_exit.is_some() {
+        asm.pad_inval_patch();
+    }
+
     // Call the branch generation function
     asm_comment!(asm, "defer_compilation");
     asm.mark_branch_start(&branch);
@@ -3307,9 +3323,10 @@ pub fn invalidate_block_version(blockref: &BlockRef) {
 
             assert!(
                 cb.get_write_ptr() <= block_end,
-                "invalidation wrote past end of block (code_size: {:?}, new_size: {})",
+                "invalidation wrote past end of block (code_size: {:?}, new_size: {}, start_addr: {:?})",
                 block.code_size(),
                 cb.get_write_ptr().as_offset() - block_start.as_offset(),
+                block.start_addr.raw_ptr(cb),
             );
             cb.set_write_ptr(cur_pos);
             cb.set_dropped_bytes(cur_dropped_bytes);

yjit/src/cruby.rs

@@ -170,6 +170,7 @@ pub use rb_iseq_encoded_size as get_iseq_encoded_size;
 pub use rb_get_iseq_body_local_iseq as get_iseq_body_local_iseq;
 pub use rb_get_iseq_body_iseq_encoded as get_iseq_body_iseq_encoded;
 pub use rb_get_iseq_body_stack_max as get_iseq_body_stack_max;
+pub use rb_get_iseq_body_type as get_iseq_body_type;
 pub use rb_get_iseq_flags_has_lead as get_iseq_flags_has_lead;
 pub use rb_get_iseq_flags_has_opt as get_iseq_flags_has_opt;
 pub use rb_get_iseq_flags_has_kw as get_iseq_flags_has_kw;

yjit/src/cruby_bindings.inc.rs

@@ -478,6 +478,16 @@ pub struct iseq_inline_iv_cache_entry {
 pub struct iseq_inline_cvar_cache_entry {
     pub entry: *mut rb_cvar_class_tbl_entry,
 }
+pub const ISEQ_TYPE_TOP: rb_iseq_type = 0;
+pub const ISEQ_TYPE_METHOD: rb_iseq_type = 1;
+pub const ISEQ_TYPE_BLOCK: rb_iseq_type = 2;
+pub const ISEQ_TYPE_CLASS: rb_iseq_type = 3;
+pub const ISEQ_TYPE_RESCUE: rb_iseq_type = 4;
+pub const ISEQ_TYPE_ENSURE: rb_iseq_type = 5;
+pub const ISEQ_TYPE_EVAL: rb_iseq_type = 6;
+pub const ISEQ_TYPE_MAIN: rb_iseq_type = 7;
+pub const ISEQ_TYPE_PLAIN: rb_iseq_type = 8;
+pub type rb_iseq_type = u32;
 pub const BUILTIN_ATTR_LEAF: rb_builtin_attr = 1;
 pub const BUILTIN_ATTR_SINGLE_NOARG_LEAF: rb_builtin_attr = 2;
 pub const BUILTIN_ATTR_INLINE_BLOCK: rb_builtin_attr = 4;
@@ -1153,6 +1163,7 @@ extern "C" {
     pub fn rb_get_iseq_body_local_table_size(iseq: *const rb_iseq_t) -> ::std::os::raw::c_uint;
     pub fn rb_get_iseq_body_iseq_encoded(iseq: *const rb_iseq_t) -> *mut VALUE;
     pub fn rb_get_iseq_body_stack_max(iseq: *const rb_iseq_t) -> ::std::os::raw::c_uint;
+    pub fn rb_get_iseq_body_type(iseq: *const rb_iseq_t) -> rb_iseq_type;
     pub fn rb_get_iseq_flags_has_lead(iseq: *const rb_iseq_t) -> bool;
     pub fn rb_get_iseq_flags_has_opt(iseq: *const rb_iseq_t) -> bool;
     pub fn rb_get_iseq_flags_has_kw(iseq: *const rb_iseq_t) -> bool;