ruby/mjit_compile.c

/**********************************************************************

  mjit_compile.c - MRI method JIT compiler

  Copyright (C) 2017 Takashi Kokubun <takashikkbn@gmail.com>.

**********************************************************************/

// NOTE: All functions in this file are executed on MJIT worker. So don't
// call Ruby methods (C functions that may call rb_funcall) or trigger
// GC (using ZALLOC, xmalloc, xfree, etc.) in this file.

#include "internal.h"

#if USE_MJIT

#include "vm_core.h"
#include "vm_exec.h"
#include "mjit.h"
#include "insns.inc"
#include "insns_info.inc"
#include "vm_insnhelper.h"

// Macros to check if a position is already compiled using compile_status.stack_size_for_pos
#define NOT_COMPILED_STACK_SIZE -1
#define ALREADY_COMPILED_P(status, pos) (status->stack_size_for_pos[pos] != NOT_COMPILED_STACK_SIZE)

// For propagating information needed for lazily pushing a frame.
struct inlined_call_context {
    int orig_argc; // ci->orig_argc
    VALUE me; // cc->me
    int param_size; // def_iseq_ptr(cc->me->def)->body->param.size
    int local_size; // def_iseq_ptr(cc->me->def)->body->local_table_size
};

// Storage to keep compiler's status.  This should have information
// which is global during one `mjit_compile` call.  Ones conditional
// in each branch should be stored in `compile_branch`.
struct compile_status {
    bool success; // has true if compilation has had no issue
    int *stack_size_for_pos; // stack_size_for_pos[pos] has stack size for the position (otherwise -1)
    // If true, JIT-ed code will use local variables to store pushed values instead of
    // using VM's stack and moving stack pointer.
    bool local_stack_p;
    // Safely-accessible cache entries copied from main thread.
    union iseq_inline_storage_entry *is_entries;
    struct rb_call_cache *cc_entries;
    // Mutated optimization levels
    struct rb_mjit_compile_info *compile_info;
    // If `inlined_iseqs[pos]` is not NULL, `mjit_compile_body` tries to inline ISeq there.
    const struct rb_iseq_constant_body **inlined_iseqs;
    struct inlined_call_context inline_context;
};

// Storage to keep data which is consistent in each conditional branch.
// This is created and used for one `compile_insns` call and its values
// should be copied for extra `compile_insns` call.
struct compile_branch {
    unsigned int stack_size; // this simulates sp (stack pointer) of YARV
    bool finish_p; // if true, compilation in this branch should stop and let another branch to be compiled
};

struct case_dispatch_var {
    FILE *f;
    unsigned int base_pos;
    VALUE last_value;
};

// Returns true if call cache is still not obsoleted and cc->me->def->type is available.
static bool
has_valid_method_type(CALL_CACHE cc)
{
    extern bool mjit_valid_class_serial_p(rb_serial_t class_serial);
    return GET_GLOBAL_METHOD_STATE() == cc->method_state
        && mjit_valid_class_serial_p(cc->class_serial) && cc->me;
}

// Returns true if iseq can use fastpath for setup, otherwise NULL. This becomes true in the same condition
// as CC_SET_FASTPATH (in vm_callee_setup_arg) is called from vm_call_iseq_setup.
static bool
fastpath_applied_iseq_p(const CALL_INFO ci, const CALL_CACHE cc, const rb_iseq_t *iseq)
{
    extern bool rb_simple_iseq_p(const rb_iseq_t *iseq);
    return iseq != NULL
        && !(ci->flag & VM_CALL_KW_SPLAT) && rb_simple_iseq_p(iseq) // Top of vm_callee_setup_arg. In this case, opt_pc is 0.
        && ci->orig_argc == iseq->body->param.lead_num // exclude argument_arity_error (assumption: `calling->argc == ci->orig_argc` in send insns)
        && vm_call_iseq_optimizable_p(ci, cc); // CC_SET_FASTPATH condition
}

static int
compile_case_dispatch_each(VALUE key, VALUE value, VALUE arg)
{
    struct case_dispatch_var *var = (struct case_dispatch_var *)arg;
    unsigned int offset;

    if (var->last_value != value) {
        offset = FIX2INT(value);
        var->last_value = value;
        fprintf(var->f, "      case %d:\n", offset);
        fprintf(var->f, "        goto label_%d;\n", var->base_pos + offset);
        fprintf(var->f, "        break;\n");
    }
    return ST_CONTINUE;
}

// Calling rb_id2str in MJIT worker causes random SEGV. So this is disabled by default.
static void
comment_id(FILE *f, ID id)
{
#ifdef MJIT_COMMENT_ID
    VALUE name = rb_id2str(id);
    const char *p, *e;
    char c, prev = '\0';

    if (!name) return;
    p = RSTRING_PTR(name);
    e = RSTRING_END(name);
    fputs("/* :\"", f);
    for (; p < e; ++p) {
        switch (c = *p) {
          case '*': case '/': if (prev != (c ^ ('/' ^ '*'))) break;
          case '\\': case '"': fputc('\\', f);
        }
        fputc(c, f);
        prev = c;
    }
    fputs("\" */", f);
#endif
}

static void compile_insns(FILE *f, const struct rb_iseq_constant_body *body, unsigned int stack_size,
                          unsigned int pos, struct compile_status *status);

// Main function of JIT compilation, vm_exec_core counterpart for JIT. Compile one insn to `f`, may modify
// b->stack_size and return next position.
//
// When you add a new instruction to insns.def, it would be nice to have JIT compilation support here but
// it's optional. This JIT compiler just ignores ISeq which includes unknown instruction, and ISeq which
// does not have it can be compiled as usual.
static unsigned int
compile_insn(FILE *f, const struct rb_iseq_constant_body *body, const int insn, const VALUE *operands,
             const unsigned int pos, struct compile_status *status, struct compile_branch *b)
{
    unsigned int next_pos = pos + insn_len(insn);

/*****************/
 #include "mjit_compile.inc"
/*****************/

    // If next_pos is already compiled and this branch is not finished yet,
    // next instruction won't be compiled in C code next and will need `goto`.
    if (!b->finish_p && next_pos < body->iseq_size && ALREADY_COMPILED_P(status, next_pos)) {
        fprintf(f, "goto label_%d;\n", next_pos);

        // Verify stack size assumption is the same among multiple branches
        if ((unsigned int)status->stack_size_for_pos[next_pos] != b->stack_size) {
            if (mjit_opts.warnings || mjit_opts.verbose)
                fprintf(stderr, "MJIT warning: JIT stack assumption is not the same between branches (%d != %u)\n",
                        status->stack_size_for_pos[next_pos], b->stack_size);
            status->success = false;
        }
    }

    return next_pos;
}

// Compile one conditional branch.  If it has branchXXX insn, this should be
// called multiple times for each branch.
static void
compile_insns(FILE *f, const struct rb_iseq_constant_body *body, unsigned int stack_size,
              unsigned int pos, struct compile_status *status)
{
    int insn;
    struct compile_branch branch;

    branch.stack_size = stack_size;
    branch.finish_p = false;

    while (pos < body->iseq_size && !ALREADY_COMPILED_P(status, pos) && !branch.finish_p) {
#if OPT_DIRECT_THREADED_CODE || OPT_CALL_THREADED_CODE
        insn = rb_vm_insn_addr2insn((void *)body->iseq_encoded[pos]);
#else
        insn = (int)body->iseq_encoded[pos];
#endif
        status->stack_size_for_pos[pos] = (int)branch.stack_size;

        fprintf(f, "\nlabel_%d: /* %s */\n", pos, insn_name(insn));
        pos = compile_insn(f, body, insn, body->iseq_encoded + (pos+1), pos, status, &branch);
        if (status->success && branch.stack_size > body->stack_max) {
            if (mjit_opts.warnings || mjit_opts.verbose)
                fprintf(stderr, "MJIT warning: JIT stack size (%d) exceeded its max size (%d)\n", branch.stack_size, body->stack_max);
            status->success = false;
        }
        if (!status->success)
            break;
    }
}

// Print the block to cancel inlined method call. It's supporting only `opt_send_without_block` for now.
static void
compile_inlined_cancel_handler(FILE *f, const struct rb_iseq_constant_body *body, struct inlined_call_context *inline_context)
{
    fprintf(f, "\ncancel:\n");
    fprintf(f, "    RB_DEBUG_COUNTER_INC(mjit_cancel);\n");
    fprintf(f, "    rb_mjit_iseq_compile_info(original_iseq->body)->disable_inlining = true;\n");
    fprintf(f, "    rb_mjit_recompile_iseq(original_iseq);\n");

    // Swap pc/sp set on cancel with original pc/sp.
    fprintf(f, "    const VALUE current_pc = reg_cfp->pc;\n");
    fprintf(f, "    const VALUE current_sp = reg_cfp->sp;\n");
    fprintf(f, "    reg_cfp->pc = orig_pc;\n");
    fprintf(f, "    reg_cfp->sp = orig_sp;\n\n");

    // Lazily push the current call frame.
    fprintf(f, "    struct rb_calling_info calling;\n");
    fprintf(f, "    calling.block_handler = VM_BLOCK_HANDLER_NONE;\n"); // assumes `opt_send_without_block`
    fprintf(f, "    calling.argc = %d;\n", inline_context->orig_argc);
    fprintf(f, "    calling.recv = reg_cfp->self;\n");
    fprintf(f, "    reg_cfp->self = orig_self;\n");
    fprintf(f, "    vm_call_iseq_setup_normal(ec, reg_cfp, &calling, (const rb_callable_method_entry_t *)0x%"PRIxVALUE", 0, %d, %d);\n\n",
            inline_context->me, inline_context->param_size, inline_context->local_size); // fastpath_applied_iseq_p checks rb_simple_iseq_p, which ensures has_opt == FALSE

    // Start usual cancel from here.
    fprintf(f, "    reg_cfp = ec->cfp;\n"); // work on the new frame
    fprintf(f, "    reg_cfp->pc = current_pc;\n");
    fprintf(f, "    reg_cfp->sp = current_sp;\n");
    for (unsigned int i = 0; i < body->stack_max; i++) { // should be always `status->local_stack_p`
        fprintf(f, "    *(vm_base_ptr(reg_cfp) + %d) = stack[%d];\n", i, i);
    }
    // We're not just returning Qundef here so that caller's normal cancel handler can
    // push back `stack` to `cfp->sp`.
    fprintf(f, "    return vm_exec(ec, ec->cfp);\n");
}

// Print the block to cancel JIT execution.
static void
compile_cancel_handler(FILE *f, const struct rb_iseq_constant_body *body, struct compile_status *status)
{
    if (status->inlined_iseqs == NULL) { // the current ISeq is being inlined
        compile_inlined_cancel_handler(f, body, &status->inline_context);
        return;
    }

    fprintf(f, "\nsend_cancel:\n");
    fprintf(f, "    RB_DEBUG_COUNTER_INC(mjit_cancel_send_inline);\n");
    fprintf(f, "    rb_mjit_iseq_compile_info(original_iseq->body)->disable_send_cache = true;\n");
    fprintf(f, "    rb_mjit_recompile_iseq(original_iseq);\n");
    fprintf(f, "    goto cancel;\n");

    fprintf(f, "\nivar_cancel:\n");
    fprintf(f, "    RB_DEBUG_COUNTER_INC(mjit_cancel_ivar_inline);\n");
    fprintf(f, "    rb_mjit_iseq_compile_info(original_iseq->body)->disable_ivar_cache = true;\n");
    fprintf(f, "    rb_mjit_recompile_iseq(original_iseq);\n");
    fprintf(f, "    goto cancel;\n");

    fprintf(f, "\ncancel:\n");
    fprintf(f, "    RB_DEBUG_COUNTER_INC(mjit_cancel);\n");
    if (status->local_stack_p) {
        for (unsigned int i = 0; i < body->stack_max; i++) {
            fprintf(f, "    *(vm_base_ptr(reg_cfp) + %d) = stack[%d];\n", i, i);
        }
    }
    fprintf(f, "    return Qundef;\n");
}

extern bool mjit_copy_cache_from_main_thread(const rb_iseq_t *iseq, struct rb_call_cache *cc_entries, union iseq_inline_storage_entry *is_entries);

static bool
mjit_compile_body(FILE *f, const rb_iseq_t *iseq, struct compile_status *status)
{
    const struct rb_iseq_constant_body *body = iseq->body;
    status->success = true;
    status->local_stack_p = !body->catch_except_p;

    if (status->local_stack_p) {
        fprintf(f, "    VALUE stack[%d];\n", body->stack_max);
    }
    else {
        fprintf(f, "    VALUE *stack = reg_cfp->sp;\n");
    }
    if (status->inlined_iseqs != NULL) // i.e. compile root
        fprintf(f, "    static const rb_iseq_t *original_iseq = 0x%"PRIxVALUE";\n", (VALUE)iseq);
    fprintf(f, "    static const VALUE *const original_body_iseq = (VALUE *)0x%"PRIxVALUE";\n",
            (VALUE)body->iseq_encoded);

    // Simulate `opt_pc` in setup_parameters_complex. Other PCs which may be passed by catch tables
    // are not considered since vm_exec doesn't call mjit_exec for catch tables.
    if (body->param.flags.has_opt) {
        int i;
        fprintf(f, "\n");
        fprintf(f, "    switch (reg_cfp->pc - reg_cfp->iseq->body->iseq_encoded) {\n");
        for (i = 0; i <= body->param.opt_num; i++) {
            VALUE pc_offset = body->param.opt_table[i];
            fprintf(f, "      case %"PRIdVALUE":\n", pc_offset);
            fprintf(f, "        goto label_%"PRIdVALUE";\n", pc_offset);
        }
        fprintf(f, "    }\n");
    }

    compile_insns(f, body, 0, 0, status);
    compile_cancel_handler(f, body, status);
    return status->success;
}

// Return true if the ISeq can be inlined without pushing a new control frame.
static bool
inlinable_iseq_p(const struct rb_iseq_constant_body *body)
{
    // 1) If catch_except_p, caller frame should be preserved when callee catches an exception.
    // Then we need to wrap `vm_exec()` but then we can't inline the call inside it.
    //
    // 2) If `body->catch_except_p` is false and `handles_sp?` of an insn is false,
    // sp is not moved as we assume `status->local_stack_p = !body->catch_except_p`.
    //
    // 3) If `body->catch_except_p` is false and `always_leaf?` of an insn is true,
    // pc is not moved.
    if (body->catch_except_p)
        return false;

    unsigned int pos = 0;
    while (pos < body->iseq_size) {
#if OPT_DIRECT_THREADED_CODE || OPT_CALL_THREADED_CODE
        int insn = rb_vm_insn_addr2insn((void *)body->iseq_encoded[pos]);
#else
        int insn = (int)body->iseq_encoded[pos];
#endif
        // All insns in the ISeq except `leave` (to be overridden in the inlined code)
        // should meet following strong assumptions:
        //   * Do not require `cfp->sp` motion
        //   * Do not move `cfp->pc`
        //   * Do not read any `cfp->pc`
        if (insn != BIN(leave) && insn_may_depend_on_sp_or_pc(insn, body->iseq_encoded + (pos + 1)))
            return false;
        pos += insn_len(insn);
    }
    return true;
}

// This needs to be macro instead of a function because it's using `alloca`.
#define INIT_COMPILE_STATUS(status, body, compile_root_p) do { \
    status = (struct compile_status){ \
        .stack_size_for_pos = (int *)alloca(sizeof(int) * body->iseq_size), \
        .inlined_iseqs = compile_root_p ? \
            alloca(sizeof(const struct rb_iseq_constant_body *) * body->iseq_size) : NULL, \
        .cc_entries = (body->ci_size + body->ci_kw_size) > 0 ? \
            alloca(sizeof(struct rb_call_cache) * (body->ci_size + body->ci_kw_size)) : NULL, \
        .is_entries = (body->is_size > 0) ? \
            alloca(sizeof(union iseq_inline_storage_entry) * body->is_size) : NULL, \
        .compile_info = compile_root_p ? \
            rb_mjit_iseq_compile_info(body) : alloca(sizeof(struct rb_mjit_compile_info)) \
    }; \
    memset(status.stack_size_for_pos, NOT_COMPILED_STACK_SIZE, sizeof(int) * body->iseq_size); \
    if (compile_root_p) \
        memset(status.inlined_iseqs, 0, sizeof(const struct rb_iseq_constant_body *) * body->iseq_size); \
    else \
        memset(status.compile_info, 0, sizeof(struct rb_mjit_compile_info)); \
} while (0)

// Compile inlinable ISeqs to C code in `f`.  It returns true if it succeeds to compile them.
static bool
precompile_inlinable_iseqs(FILE *f, const rb_iseq_t *iseq, struct compile_status *status)
{
    const struct rb_iseq_constant_body *body = iseq->body;
    unsigned int pos = 0;
    while (pos < body->iseq_size) {
#if OPT_DIRECT_THREADED_CODE || OPT_CALL_THREADED_CODE
        int insn = rb_vm_insn_addr2insn((void *)body->iseq_encoded[pos]);
#else
        int insn = (int)body->iseq_encoded[pos];
#endif

        if (insn == BIN(opt_send_without_block)) { // `compile_inlined_cancel_handler` supports only `opt_send_without_block`
            CALL_INFO ci = (CALL_INFO)body->iseq_encoded[pos + 1];
            CALL_CACHE cc_copy = status->cc_entries + ((CALL_CACHE)body->iseq_encoded[pos + 2] - body->cc_entries); // use copy to avoid race condition

            const rb_iseq_t *child_iseq;
            if (has_valid_method_type(cc_copy) &&
                    !(ci->flag & VM_CALL_TAILCALL) && // inlining only non-tailcall path
                    cc_copy->me->def->type == VM_METHOD_TYPE_ISEQ && fastpath_applied_iseq_p(ci, cc_copy, child_iseq = def_iseq_ptr(cc_copy->me->def)) && // CC_SET_FASTPATH in vm_callee_setup_arg
                    inlinable_iseq_p(child_iseq->body)) {
                status->inlined_iseqs[pos] = child_iseq->body;

                if (mjit_opts.verbose >= 1) // print beforehand because ISeq may be GCed during copy job.
                    fprintf(stderr, "JIT inline: %s@%s:%d => %s@%s:%d\n",
                            RSTRING_PTR(iseq->body->location.label),
                            RSTRING_PTR(rb_iseq_path(iseq)), FIX2INT(iseq->body->location.first_lineno),
                            RSTRING_PTR(child_iseq->body->location.label),
                            RSTRING_PTR(rb_iseq_path(child_iseq)), FIX2INT(child_iseq->body->location.first_lineno));

                struct compile_status child_status;
                INIT_COMPILE_STATUS(child_status, child_iseq->body, false);
                child_status.inline_context = (struct inlined_call_context){
                    .orig_argc = ci->orig_argc,
                    .me = (VALUE)cc_copy->me,
                    .param_size = child_iseq->body->param.size,
                    .local_size = child_iseq->body->local_table_size
                };
                if ((child_status.cc_entries != NULL || child_status.is_entries != NULL)
                        && !mjit_copy_cache_from_main_thread(child_iseq, child_status.cc_entries, child_status.is_entries))
                    return false;

                fprintf(f, "ALWAYS_INLINE(static VALUE _mjit_inlined_%d(rb_execution_context_t *ec, rb_control_frame_t *reg_cfp, const VALUE orig_self, const rb_iseq_t *original_iseq));\n", pos);
                fprintf(f, "static inline VALUE\n_mjit_inlined_%d(rb_execution_context_t *ec, rb_control_frame_t *reg_cfp, const VALUE orig_self, const rb_iseq_t *original_iseq)\n{\n", pos);
                fprintf(f, "    const VALUE *orig_pc = reg_cfp->pc;\n");
                fprintf(f, "    const VALUE *orig_sp = reg_cfp->sp;\n");
                bool success = mjit_compile_body(f, child_iseq, &child_status);
                fprintf(f, "\n} /* end of _mjit_inlined_%d */\n\n", pos);

                if (!success)
                    return false;
            }
        }
        pos += insn_len(insn);
    }
    return true;
}

// Compile ISeq to C code in `f`. It returns true if it succeeds to compile.
bool
mjit_compile(FILE *f, const rb_iseq_t *iseq, const char *funcname)
{
    // For performance, we verify stack size only on compilation time (mjit_compile.inc.erb) without --jit-debug
    if (!mjit_opts.debug) {
        fprintf(f, "#undef OPT_CHECKED_RUN\n");
        fprintf(f, "#define OPT_CHECKED_RUN 0\n\n");
    }

    struct compile_status status;
    INIT_COMPILE_STATUS(status, iseq->body, true);
    if ((status.cc_entries != NULL || status.is_entries != NULL)
            && !mjit_copy_cache_from_main_thread(iseq, status.cc_entries, status.is_entries))
        return false;

    if (!status.compile_info->disable_send_cache && !status.compile_info->disable_inlining) {
        if (!precompile_inlinable_iseqs(f, iseq, &status))
            return false;
    }

#ifdef _WIN32
    fprintf(f, "__declspec(dllexport)\n");
#endif
    fprintf(f, "VALUE\n%s(rb_execution_context_t *ec, rb_control_frame_t *reg_cfp)\n{\n", funcname);
    bool success = mjit_compile_body(f, iseq, &status);
    fprintf(f, "\n} // end of %s\n", funcname);
    return success;
}

#endif // USE_MJIT