view release on metacpan or  search on metacpan

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

** an any/any rule to avoid being passed on to CSE.
**
** Carefully read the following requirements before adding or modifying
** any fold rules:
**
** Requirement #1: All fold rules must preserve their destination type.
**
** Consistently use INTFOLD() (KINT result) or lj_ir_knum() (KNUM result).
** Never use lj_ir_knumint() which can have either a KINT or KNUM result.
**
** Requirement #2: Fold rules should not create *new* instructions which
** reference operands *across* PHIs.
**
** E.g. a RETRYFOLD with 'fins->op1 = fleft->op1' is invalid if the
** left operand is a PHI. Then fleft->op1 would point across the PHI
** frontier to an invariant instruction. Adding a PHI for this instruction
** would be counterproductive. The solution is to add a barrier which
** prevents folding across PHIs, i.e. 'PHIBARRIER(fleft)' in this case.
** The only exception is for recurrences with high latencies like
** repeated int->num->int conversions.
**
** One could relax this condition a bit if the referenced instruction is
** a PHI, too. But this often leads to worse code due to excessive
** register shuffling.
**
** Note: returning *existing* instructions (e.g. LEFTFOLD) is ok, though.
** Even returning fleft->op1 would be ok, because a new PHI will added,
** if needed. But again, this leads to excessive register shuffling and
** should be avoided.
**
** Requirement #3: The set of all fold rules must be monotonic to guarantee
** termination.
**
** The goal is optimization, so one primarily wants to add strength-reducing
** rules. This means eliminating an instruction or replacing an instruction
** with one or more simpler instructions. Don't add fold rules which point
** into the other direction.
**
** Some rules (like commutativity) do not directly reduce the strength of
** an instruction, but enable other fold rules (e.g. by moving constants
** to the right operand). These rules must be made unidirectional to avoid
** cycles.
**
** Rule of thumb: the trace recorder expands the IR and FOLD shrinks it.
*/

/* Some local macros to save typing. Undef'd at the end. */
#define IR(ref)		(&J->cur.ir[(ref)])
#define fins		(&J->fold.ins)
#define fleft		(J->fold.left)
#define fright		(J->fold.right)
#define knumleft	(ir_knum(fleft)->n)
#define knumright	(ir_knum(fright)->n)

/* 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))

/* Fold function type. Fastcall on x86 significantly reduces their size. */
typedef IRRef (LJ_FASTCALL *FoldFunc)(jit_State *J);

/* Macros for the fold specs, so buildvm can recognize them. */
#define LJFOLD(x)
#define LJFOLDX(x)
#define LJFOLDF(name)	static TRef LJ_FASTCALL fold_##name(jit_State *J)
/* Note: They must be at the start of a line or buildvm ignores them! */

/* Barrier to prevent using operands across PHIs. */
#define PHIBARRIER(ir)	if (irt_isphi((ir)->t)) return NEXTFOLD

/* Barrier to prevent folding across a GC step.
** GC steps can only happen at the head of a trace and at LOOP.
** And the GC is only driven forward if there's at least one allocation.
*/
#define gcstep_barrier(J, ref) \
  ((ref) < J->chain[IR_LOOP] && \
   (J->chain[IR_SNEW] || J->chain[IR_XSNEW] || \
    J->chain[IR_TNEW] || J->chain[IR_TDUP] || \
    J->chain[IR_CNEW] || J->chain[IR_CNEWI] || \
    J->chain[IR_BUFSTR] || J->chain[IR_TOSTR] || J->chain[IR_CALLA]))

/* -- Constant folding for FP numbers ------------------------------------- */

LJFOLD(ADD KNUM KNUM)
LJFOLD(SUB KNUM KNUM)
LJFOLD(MUL KNUM KNUM)
LJFOLD(DIV KNUM KNUM)
LJFOLD(LDEXP KNUM KNUM)
LJFOLD(MIN KNUM KNUM)
LJFOLD(MAX KNUM KNUM)
LJFOLDF(kfold_numarith)
{
  lua_Number a = knumleft;
  lua_Number b = knumright;
  lua_Number y = lj_vm_foldarith(a, b, fins->o - IR_ADD);
  return lj_ir_knum(J, y);
}

LJFOLD(NEG KNUM FLOAD)
LJFOLD(ABS KNUM FLOAD)
LJFOLDF(kfold_numabsneg)
{
  lua_Number a = knumleft;
  lua_Number y = lj_vm_foldarith(a, a, fins->o - IR_ADD);
  return lj_ir_knum(J, y);
}

LJFOLD(LDEXP KNUM KINT)
LJFOLDF(kfold_ldexp)
{
#if LJ_TARGET_X86ORX64
  UNUSED(J);
  return NEXTFOLD;
#else
  return lj_ir_knum(J, ldexp(knumleft, fright->i));
#endif
}

LJFOLD(FPMATH KNUM any)
LJFOLDF(kfold_fpmath)
{
  lua_Number a = knumleft;