view release on metacpan or  search on metacpan

ffi/target/release/build/mlua-sys-6a99a2ae50f12319/out/luajit-build/build/src/lj_opt_loop.c  view on Meta::CPAN

/*
** LOOP: Loop Optimizations.
** Copyright (C) 2005-2025 Mike Pall. See Copyright Notice in luajit.h
*/

#define lj_opt_loop_c
#define LUA_CORE

#include "lj_obj.h"

#if LJ_HASJIT

#include "lj_err.h"
#include "lj_buf.h"
#include "lj_ir.h"
#include "lj_jit.h"
#include "lj_iropt.h"
#include "lj_trace.h"
#include "lj_snap.h"
#include "lj_vm.h"

/* Loop optimization:
**
** Traditional Loop-Invariant Code Motion (LICM) splits the instructions
** of a loop into invariant and variant instructions. The invariant
** instructions are hoisted out of the loop and only the variant
** instructions remain inside the loop body.
**
** Unfortunately LICM is mostly useless for compiling dynamic languages.
** The IR has many guards and most of the subsequent instructions are
** control-dependent on them. The first non-hoistable guard would
** effectively prevent hoisting of all subsequent instructions.
**
** That's why we use a special form of unrolling using copy-substitution,
** combined with redundancy elimination:
**
** The recorded instruction stream is re-emitted to the compiler pipeline
** with substituted operands. The substitution table is filled with the
** refs returned by re-emitting each instruction. This can be done
** on-the-fly, because the IR is in strict SSA form, where every ref is
** defined before its use.
**
** This aproach generates two code sections, separated by the LOOP
** instruction:
**
** 1. The recorded instructions form a kind of pre-roll for the loop. It
** contains a mix of invariant and variant instructions and performs
** exactly one loop iteration (but not necessarily the 1st iteration).
**
** 2. The loop body contains only the variant instructions and performs
** all remaining loop iterations.
**
** On first sight that looks like a waste of space, because the variant
** instructions are present twice. But the key insight is that the
** pre-roll honors the control-dependencies for *both* the pre-roll itself
** *and* the loop body!
**
** It also means one doesn't have to explicitly model control-dependencies
** (which, BTW, wouldn't help LICM much). And it's much easier to
** integrate sparse snapshotting with this approach.
**
** One of the nicest aspects of this approach is that all of the
** optimizations of the compiler pipeline (FOLD, CSE, FWD, etc.) can be
** reused with only minor restrictions (e.g. one should not fold
** instructions across loop-carried dependencies).
**
** But in general all optimizations can be applied which only need to look
** backwards into the generated instruction stream. At any point in time
** during the copy-substitution process this contains both a static loop
** iteration (the pre-roll) and a dynamic one (from the to-be-copied
** instruction up to the end of the partial loop body).
**
** Since control-dependencies are implicitly kept, CSE also applies to all
** kinds of guards. The major advantage is that all invariant guards can
** be hoisted, too.
**
** Load/store forwarding works across loop iterations, too. This is
** important if loop-carried dependencies are kept in upvalues or tables.
** E.g. 'self.idx = self.idx + 1' deep down in some OO-style method may
** become a forwarded loop-recurrence after inlining.
**
** Since the IR is in SSA form, loop-carried dependencies have to be
** modeled with PHI instructions. The potential candidates for PHIs are
** collected on-the-fly during copy-substitution. After eliminating the
** redundant ones, PHI instructions are emitted *below* the loop body.
**
** Note that this departure from traditional SSA form doesn't change the
** semantics of the PHI instructions themselves. But it greatly simplifies
** on-the-fly generation of the IR and the machine code.
*/

/* Some local macros to save typing. Undef'd at the end. */
#define IR(ref)		(&J->cur.ir[(ref)])

/* Pass IR on to next optimization in chain (FOLD). */
#define emitir(ot, a, b)	(lj_ir_set(J, (ot), (a), (b)), lj_opt_fold(J))

/* Emit raw IR without passing through optimizations. */
#define emitir_raw(ot, a, b)	(lj_ir_set(J, (ot), (a), (b)), lj_ir_emit(J))

/* -- PHI elimination ----------------------------------------------------- */

/* Emit or eliminate collected PHIs. */
static void loop_emit_phi(jit_State *J, IRRef1 *subst, IRRef1 *phi, IRRef nphi,
			  SnapNo onsnap)
{
  int passx = 0;
  IRRef i, j, nslots;
  IRRef invar = J->chain[IR_LOOP];
  /* Pass #1: mark redundant and potentially redundant PHIs. */
  for (i = 0, j = 0; i < nphi; i++) {
    IRRef lref = phi[i];
    IRRef rref = subst[lref];
    if (lref == rref || rref == REF_DROP) {  /* Invariants are redundant. */
      irt_clearphi(IR(lref)->t);
    } else {
      phi[j++] = (IRRef1)lref;
      if (!(IR(rref)->op1 == lref || IR(rref)->op2 == lref)) {
	/* Quick check for simple recurrences failed, need pass2. */
	irt_setmark(IR(lref)->t);
	passx = 1;
      }
    }
  }
  nphi = j;
  /* Pass #2: traverse variant part and clear marks of non-redundant PHIs. */
  if (passx) {
    SnapNo s;
    for (i = J->cur.nins-1; i > invar; i--) {
      IRIns *ir = IR(i);
      if (!irref_isk(ir->op2)) irt_clearmark(IR(ir->op2)->t);
      if (!irref_isk(ir->op1)) {
	irt_clearmark(IR(ir->op1)->t);
	if (ir->op1 < invar &&
	    ir->o >= IR_CALLN && ir->o <= IR_CARG) {  /* ORDER IR */
	  ir = IR(ir->op1);
	  while (ir->o == IR_CARG) {