ruby--ruby/tool/ruby_vm/views/_mjit_compile_send.erb

% # -*- C -*-
% # Copyright (c) 2018 Takashi Kokubun.  All rights reserved.
% #
% # This file is a part of  the programming language Ruby.  Permission is hereby
% # granted, to either  redistribute and/or modify this file,  provided that the
% # conditions mentioned  in the  file COPYING  are met.   Consult the  file for
% # details.
%
% # Optimized case of send / opt_send_without_block instructions.
{
% # compiler: Prepare operands which may be used by `insn.call_attribute`
% insn.opes.each_with_index do |ope, i|
    MAYBE_UNUSED(<%= ope.fetch(:decl) %>) = (<%= ope.fetch(:type) %>)operands[<%= i %>];
% end
% # compiler: Use captured cc to avoid race condition
    const struct rb_callcache *captured_cc = captured_cc_entries(status)[call_data_index(cd, body)];
%
% # compiler: Inline send insn where some supported fastpath is used.
    const rb_iseq_t *iseq = NULL;
    const CALL_INFO ci = cd->ci;
    int kw_splat = IS_ARGS_KW_SPLAT(ci) > 0;
    extern bool rb_splat_or_kwargs_p(const struct rb_callinfo *restrict ci);
    if (!status->compile_info->disable_send_cache && has_valid_method_type(captured_cc) && (
%       # `CC_SET_FASTPATH(cd->cc, vm_call_cfunc_with_frame, ...)` in `vm_call_cfunc`
        (vm_cc_cme(captured_cc)->def->type == VM_METHOD_TYPE_CFUNC
         && !rb_splat_or_kwargs_p(ci) && !kw_splat)
%       # `CC_SET_FASTPATH(cc, vm_call_iseq_setup_func(...), vm_call_iseq_optimizable_p(...))` in `vm_callee_setup_arg`,
%       # and support only non-VM_CALL_TAILCALL path inside it
        || (vm_cc_cme(captured_cc)->def->type == VM_METHOD_TYPE_ISEQ
            && fastpath_applied_iseq_p(ci, captured_cc, iseq = def_iseq_ptr(vm_cc_cme(captured_cc)->def))
            && !(vm_ci_flag(ci) & VM_CALL_TAILCALL))
    )) {
        const bool cfunc_debug = false; // Set true when you want to see inlined cfunc
        if (cfunc_debug && vm_cc_cme(captured_cc)->def->type == VM_METHOD_TYPE_CFUNC)
            fprintf(stderr, "  * %s\n", rb_id2name(vm_ci_mid(ci)));

        int sp_inc = (int)sp_inc_of_sendish(ci);
        fprintf(f, "{\n");

% # JIT: Invalidate call cache if it requires vm_search_method. This allows to inline some of following things.
        bool opt_class_of = !maybe_special_const_class_p(captured_cc->klass); // If true, use RBASIC_CLASS instead of CLASS_OF to reduce code size
        fprintf(f, "    const struct rb_callcache *cc = (const struct rb_callcache *)0x%"PRIxVALUE";\n", (VALUE)captured_cc);
        fprintf(f, "    const rb_callable_method_entry_t *cc_cme = (const rb_callable_method_entry_t *)0x%"PRIxVALUE";\n", (VALUE)vm_cc_cme(captured_cc));
        fprintf(f, "    const VALUE recv = stack[%d];\n", b->stack_size + sp_inc - 1);
        fprintf(f, "    if (UNLIKELY(%s || !vm_cc_valid_p(cc, cc_cme, %s(recv)))) {\n", opt_class_of ? "RB_SPECIAL_CONST_P(recv)" : "false", opt_class_of ? "RBASIC_CLASS" : "CLASS_OF");
        fprintf(f, "        reg_cfp->pc = original_body_iseq + %d;\n", pos);
        fprintf(f, "        reg_cfp->sp = vm_base_ptr(reg_cfp) + %d;\n", b->stack_size);
        fprintf(f, "        goto send_cancel;\n");
        fprintf(f, "    }\n");

% # JIT: move sp and pc if necessary
<%= render 'mjit_compile_pc_and_sp', locals: { insn: insn } -%>

% # JIT: If ISeq is inlinable, call the inlined method without pushing a frame.
        if (iseq && status->inlined_iseqs != NULL && iseq->body == status->inlined_iseqs[pos]) {
            fprintf(f, "    {\n");
            fprintf(f, "        VALUE orig_self = reg_cfp->self;\n");
            fprintf(f, "        reg_cfp->self = stack[%d];\n", b->stack_size + sp_inc - 1);
            fprintf(f, "        stack[%d] = _mjit%d_inlined_%d(ec, reg_cfp, orig_self, original_iseq);\n", b->stack_size + sp_inc - 1, status->compiled_id, pos);
            fprintf(f, "        reg_cfp->self = orig_self;\n");
            fprintf(f, "    }\n");
        }
        else {
% # JIT: Forked `vm_sendish` (except method_explorer = vm_search_method_wrap) to inline various things
            fprintf(f, "    {\n");
            fprintf(f, "        VALUE val;\n");
            fprintf(f, "        struct rb_calling_info calling;\n");
% if insn.name == 'send'
            fprintf(f, "        calling.block_handler = vm_caller_setup_arg_block(ec, reg_cfp, (const struct rb_callinfo *)0x%"PRIxVALUE", (rb_iseq_t *)0x%"PRIxVALUE", FALSE);\n", (VALUE)ci, (VALUE)blockiseq);
% else
            fprintf(f, "        calling.block_handler = VM_BLOCK_HANDLER_NONE;\n");
% end
            fprintf(f, "        calling.kw_splat = %d;\n", kw_splat);
            fprintf(f, "        calling.recv = stack[%d];\n", b->stack_size + sp_inc - 1);
            fprintf(f, "        calling.argc = %d;\n", vm_ci_argc(ci));

            if (vm_cc_cme(captured_cc)->def->type == VM_METHOD_TYPE_CFUNC) {
%               # TODO: optimize this more
                fprintf(f, "        struct rb_call_data cc_cd = { .ci = (CALL_INFO)0x%"PRIxVALUE", .cc = cc };\n", (VALUE)ci); // creating local cd here because operand's cd->cc may not be the same as inlined cc.
                fprintf(f, "        val = vm_call_cfunc_with_frame(ec, reg_cfp, &calling, &cc_cd);\n");
            }
            else { // VM_METHOD_TYPE_ISEQ
%               # fastpath_applied_iseq_p checks rb_simple_iseq_p, which ensures has_opt == FALSE
                fprintf(f, "        vm_call_iseq_setup_normal(ec, reg_cfp, &calling, cc_cme, 0, %d, %d);\n", iseq->body->param.size, iseq->body->local_table_size);
                if (iseq->body->catch_except_p) {
                    fprintf(f, "        VM_ENV_FLAGS_SET(ec->cfp->ep, VM_FRAME_FLAG_FINISH);\n");
                    fprintf(f, "        val = vm_exec(ec, TRUE);\n");
                }
                else {
                    fprintf(f, "        if ((val = mjit_exec(ec)) == Qundef) {\n");
                    fprintf(f, "            VM_ENV_FLAGS_SET(ec->cfp->ep, VM_FRAME_FLAG_FINISH);\n"); // This is vm_call0_body's code after vm_call_iseq_setup
                    fprintf(f, "            val = vm_exec(ec, FALSE);\n");
                    fprintf(f, "        }\n");
                }
            }
            fprintf(f, "        stack[%d] = val;\n", b->stack_size + sp_inc - 1);
            fprintf(f, "    }\n");

% # JIT: We should evaluate ISeq modified for TracePoint if it's enabled. Note: This is slow.
            fprintf(f, "    if (UNLIKELY(!mjit_call_p)) {\n");
            fprintf(f, "        reg_cfp->sp = vm_base_ptr(reg_cfp) + %d;\n", b->stack_size + (int)<%= insn.call_attribute('sp_inc') %>);
            if (!pc_moved_p) {
                fprintf(f, "        reg_cfp->pc = original_body_iseq + %d;\n", next_pos);
            }
            fprintf(f, "        RB_DEBUG_COUNTER_INC(mjit_cancel_invalidate_all);\n");
            fprintf(f, "        goto cancel;\n");
            fprintf(f, "    }\n");
        }

% # compiler: Move JIT compiler's internal stack pointer
        b->stack_size += <%= insn.call_attribute('sp_inc') %>;

        fprintf(f, "}\n");
        break;
    }
}