2 ** Memory access optimizations.
3 ** AA: Alias Analysis using high-level semantic disambiguation.
4 ** FWD: Load Forwarding (L2L) + Store Forwarding (S2L).
5 ** DSE: Dead-Store Elimination.
6 ** Copyright (C) 2005-2012 Mike Pall. See Copyright Notice in luajit.h
21 /* Some local macros to save typing. Undef'd at the end. */
22 #define IR(ref) (&J->cur.ir[(ref)])
23 #define fins (&J->fold.ins)
24 #define fleft (&J->fold.left)
25 #define fright (&J->fold.right)
28 ** Caveat #1: return value is not always a TRef -- only use with tref_ref().
29 ** Caveat #2: FWD relies on active CSE for xREF operands -- see lj_opt_fold().
32 /* Return values from alias analysis. */
34 ALIAS_NO
, /* The two refs CANNOT alias (exact). */
35 ALIAS_MAY
, /* The two refs MAY alias (inexact). */
36 ALIAS_MUST
/* The two refs MUST alias (exact). */
39 /* -- ALOAD/HLOAD forwarding and ASTORE/HSTORE elimination ---------------- */
41 /* Simplified escape analysis: check for intervening stores. */
42 static AliasRet
aa_escape(jit_State
*J
, IRIns
*ir
, IRIns
*stop
)
44 IRRef ref
= (IRRef
)(ir
- J
->cur
.ir
); /* The ref that might be stored. */
45 for (ir
++; ir
< stop
; ir
++)
47 (ir
->o
== IR_ASTORE
|| ir
->o
== IR_HSTORE
||
48 ir
->o
== IR_USTORE
|| ir
->o
== IR_FSTORE
))
49 return ALIAS_MAY
; /* Reference was stored and might alias. */
50 return ALIAS_NO
; /* Reference was not stored. */
53 /* Alias analysis for two different table references. */
54 static AliasRet
aa_table(jit_State
*J
, IRRef ta
, IRRef tb
)
56 IRIns
*taba
= IR(ta
), *tabb
= IR(tb
);
59 lua_assert(irt_istab(taba
->t
) && irt_istab(tabb
->t
));
60 /* Disambiguate new allocations. */
61 newa
= (taba
->o
== IR_TNEW
|| taba
->o
== IR_TDUP
);
62 newb
= (tabb
->o
== IR_TNEW
|| tabb
->o
== IR_TDUP
);
64 return ALIAS_NO
; /* Two different allocations never alias. */
65 if (newb
) { /* At least one allocation? */
66 IRIns
*tmp
= taba
; taba
= tabb
; tabb
= tmp
;
68 return ALIAS_MAY
; /* Anything else: we just don't know. */
70 return aa_escape(J
, taba
, tabb
);
73 /* Alias analysis for array and hash access using key-based disambiguation. */
74 static AliasRet
aa_ahref(jit_State
*J
, IRIns
*refa
, IRIns
*refb
)
81 return ALIAS_MUST
; /* Shortcut for same refs. */
83 if (keya
->o
== IR_KSLOT
) { ka
= keya
->op1
; keya
= IR(ka
); }
85 if (keyb
->o
== IR_KSLOT
) { kb
= keyb
->op1
; keyb
= IR(kb
); }
86 ta
= (refa
->o
==IR_HREFK
|| refa
->o
==IR_AREF
) ? IR(refa
->op1
)->op1
: refa
->op1
;
87 tb
= (refb
->o
==IR_HREFK
|| refb
->o
==IR_AREF
) ? IR(refb
->op1
)->op1
: refb
->op1
;
89 /* Same key. Check for same table with different ref (NEWREF vs. HREF). */
91 return ALIAS_MUST
; /* Same key, same table. */
93 return aa_table(J
, ta
, tb
); /* Same key, possibly different table. */
95 if (irref_isk(ka
) && irref_isk(kb
))
96 return ALIAS_NO
; /* Different constant keys. */
97 if (refa
->o
== IR_AREF
) {
98 /* Disambiguate array references based on index arithmetic. */
99 int32_t ofsa
= 0, ofsb
= 0;
100 IRRef basea
= ka
, baseb
= kb
;
101 lua_assert(refb
->o
== IR_AREF
);
102 /* Gather base and offset from t[base] or t[base+-ofs]. */
103 if (keya
->o
== IR_ADD
&& irref_isk(keya
->op2
)) {
105 ofsa
= IR(keya
->op2
)->i
;
106 if (basea
== kb
&& ofsa
!= 0)
107 return ALIAS_NO
; /* t[base+-ofs] vs. t[base]. */
109 if (keyb
->o
== IR_ADD
&& irref_isk(keyb
->op2
)) {
111 ofsb
= IR(keyb
->op2
)->i
;
112 if (ka
== baseb
&& ofsb
!= 0)
113 return ALIAS_NO
; /* t[base] vs. t[base+-ofs]. */
115 if (basea
== baseb
&& ofsa
!= ofsb
)
116 return ALIAS_NO
; /* t[base+-o1] vs. t[base+-o2] and o1 != o2. */
118 /* Disambiguate hash references based on the type of their keys. */
119 lua_assert((refa
->o
==IR_HREF
|| refa
->o
==IR_HREFK
|| refa
->o
==IR_NEWREF
) &&
120 (refb
->o
==IR_HREF
|| refb
->o
==IR_HREFK
|| refb
->o
==IR_NEWREF
));
121 if (!irt_sametype(keya
->t
, keyb
->t
))
122 return ALIAS_NO
; /* Different key types. */
125 return ALIAS_MAY
; /* Same table, cannot disambiguate keys. */
127 return aa_table(J
, ta
, tb
); /* Try to disambiguate tables. */
130 /* Array and hash load forwarding. */
131 static TRef
fwd_ahload(jit_State
*J
, IRRef xref
)
133 IRIns
*xr
= IR(xref
);
134 IRRef lim
= xref
; /* Search limit. */
137 /* Search for conflicting stores. */
138 ref
= J
->chain
[fins
->o
+IRDELTA_L2S
];
140 IRIns
*store
= IR(ref
);
141 switch (aa_ahref(J
, xr
, IR(store
->op1
))) {
142 case ALIAS_NO
: break; /* Continue searching. */
143 case ALIAS_MAY
: lim
= ref
; goto cselim
; /* Limit search for load. */
144 case ALIAS_MUST
: return store
->op2
; /* Store forwarding. */
149 /* No conflicting store (yet): const-fold loads from allocations. */
151 IRIns
*ir
= (xr
->o
== IR_HREFK
|| xr
->o
== IR_AREF
) ? IR(xr
->op1
) : xr
;
154 if (ir
->o
== IR_TNEW
|| (ir
->o
== IR_TDUP
&& irref_isk(xr
->op2
))) {
155 /* A NEWREF with a number key may end up pointing to the array part.
156 ** But it's referenced from HSTORE and not found in the ASTORE chain.
157 ** For now simply consider this a conflict without forwarding anything.
159 if (xr
->o
== IR_AREF
) {
160 IRRef ref2
= J
->chain
[IR_NEWREF
];
162 IRIns
*newref
= IR(ref2
);
163 if (irt_isnum(IR(newref
->op2
)->t
))
168 /* NEWREF inhibits CSE for HREF, and dependent FLOADs from HREFK/AREF.
169 ** But the above search for conflicting stores was limited by xref.
170 ** So continue searching, limited by the TNEW/TDUP. Store forwarding
171 ** is ok, too. A conflict does NOT limit the search for a matching load.
174 IRIns
*store
= IR(ref
);
175 switch (aa_ahref(J
, xr
, IR(store
->op1
))) {
176 case ALIAS_NO
: break; /* Continue searching. */
177 case ALIAS_MAY
: goto cselim
; /* Conflicting store. */
178 case ALIAS_MUST
: return store
->op2
; /* Store forwarding. */
182 lua_assert(ir
->o
!= IR_TNEW
|| irt_isnil(fins
->t
));
183 if (irt_ispri(fins
->t
)) {
184 return TREF_PRI(irt_type(fins
->t
));
185 } else if (irt_isnum(fins
->t
) || (LJ_DUALNUM
&& irt_isint(fins
->t
)) ||
186 irt_isstr(fins
->t
)) {
189 IRIns
*key
= IR(xr
->op2
);
190 if (key
->o
== IR_KSLOT
) key
= IR(key
->op1
);
191 lj_ir_kvalue(J
->L
, &keyv
, key
);
192 tv
= lj_tab_get(J
->L
, ir_ktab(IR(ir
->op1
)), &keyv
);
193 lua_assert(itype2irt(tv
) == irt_type(fins
->t
));
194 if (irt_isnum(fins
->t
))
195 return lj_ir_knum_u64(J
, tv
->u64
);
196 else if (LJ_DUALNUM
&& irt_isint(fins
->t
))
197 return lj_ir_kint(J
, intV(tv
));
199 return lj_ir_kstr(J
, strV(tv
));
201 /* Othwerwise: don't intern as a constant. */
206 /* Try to find a matching load. Below the conflicting store, if any. */
207 ref
= J
->chain
[fins
->o
];
209 IRIns
*load
= IR(ref
);
210 if (load
->op1
== xref
)
211 return ref
; /* Load forwarding. */
214 return 0; /* Conflict or no match. */
217 /* Reassociate ALOAD across PHIs to handle t[i-1] forwarding case. */
218 static TRef
fwd_aload_reassoc(jit_State
*J
)
220 IRIns
*irx
= IR(fins
->op1
);
221 IRIns
*key
= IR(irx
->op2
);
222 if (key
->o
== IR_ADD
&& irref_isk(key
->op2
)) {
223 IRIns
*add2
= IR(key
->op1
);
224 if (add2
->o
== IR_ADD
&& irref_isk(add2
->op2
) &&
225 IR(key
->op2
)->i
== -IR(add2
->op2
)->i
) {
226 IRRef ref
= J
->chain
[IR_AREF
];
227 IRRef lim
= add2
->op1
;
228 if (irx
->op1
> lim
) lim
= irx
->op1
;
231 if (ir
->op1
== irx
->op1
&& ir
->op2
== add2
->op1
)
232 return fwd_ahload(J
, ref
);
240 /* ALOAD forwarding. */
241 TRef LJ_FASTCALL
lj_opt_fwd_aload(jit_State
*J
)
244 if ((ref
= fwd_ahload(J
, fins
->op1
)) ||
245 (ref
= fwd_aload_reassoc(J
)))
250 /* HLOAD forwarding. */
251 TRef LJ_FASTCALL
lj_opt_fwd_hload(jit_State
*J
)
253 IRRef ref
= fwd_ahload(J
, fins
->op1
);
259 /* HREFK forwarding. */
260 TRef LJ_FASTCALL
lj_opt_fwd_hrefk(jit_State
*J
)
262 IRRef tab
= fleft
->op1
;
263 IRRef ref
= J
->chain
[IR_NEWREF
];
265 IRIns
*newref
= IR(ref
);
266 if (tab
== newref
->op1
) {
267 if (fright
->op1
== newref
->op2
)
268 return ref
; /* Forward from NEWREF. */
271 } else if (aa_table(J
, tab
, newref
->op1
) != ALIAS_NO
) {
276 /* No conflicting NEWREF: key location unchanged for HREFK of TDUP. */
277 if (IR(tab
)->o
== IR_TDUP
)
278 fins
->t
.irt
&= ~IRT_GUARD
; /* Drop HREFK guard. */
283 /* Check whether HREF of TNEW/TDUP can be folded to niltv. */
284 int LJ_FASTCALL
lj_opt_fwd_href_nokey(jit_State
*J
)
286 IRRef lim
= fins
->op1
; /* Search limit. */
289 /* The key for an ASTORE may end up in the hash part after a NEWREF. */
290 if (irt_isnum(fright
->t
) && J
->chain
[IR_NEWREF
] > lim
) {
291 ref
= J
->chain
[IR_ASTORE
];
293 if (ref
< J
->chain
[IR_NEWREF
])
294 return 0; /* Conflict. */
299 /* Search for conflicting stores. */
300 ref
= J
->chain
[IR_HSTORE
];
302 IRIns
*store
= IR(ref
);
303 if (aa_ahref(J
, fins
, IR(store
->op1
)) != ALIAS_NO
)
304 return 0; /* Conflict. */
308 return 1; /* No conflict. Can fold to niltv. */
311 /* Check whether there's no aliasing NEWREF for the left operand. */
312 int LJ_FASTCALL
lj_opt_fwd_tptr(jit_State
*J
, IRRef lim
)
314 IRRef ta
= fins
->op1
;
315 IRRef ref
= J
->chain
[IR_NEWREF
];
317 IRIns
*newref
= IR(ref
);
318 if (ta
== newref
->op1
|| aa_table(J
, ta
, newref
->op1
) != ALIAS_NO
)
319 return 0; /* Conflict. */
322 return 1; /* No conflict. Can safely FOLD/CSE. */
325 /* ASTORE/HSTORE elimination. */
326 TRef LJ_FASTCALL
lj_opt_dse_ahstore(jit_State
*J
)
328 IRRef xref
= fins
->op1
; /* xREF reference. */
329 IRRef val
= fins
->op2
; /* Stored value reference. */
330 IRIns
*xr
= IR(xref
);
331 IRRef1
*refp
= &J
->chain
[fins
->o
];
333 while (ref
> xref
) { /* Search for redundant or conflicting stores. */
334 IRIns
*store
= IR(ref
);
335 switch (aa_ahref(J
, xr
, IR(store
->op1
))) {
337 break; /* Continue searching. */
338 case ALIAS_MAY
: /* Store to MAYBE the same location. */
339 if (store
->op2
!= val
) /* Conflict if the value is different. */
341 break; /* Otherwise continue searching. */
342 case ALIAS_MUST
: /* Store to the same location. */
343 if (store
->op2
== val
) /* Same value: drop the new store. */
345 /* Different value: try to eliminate the redundant store. */
346 if (ref
> J
->chain
[IR_LOOP
]) { /* Quick check to avoid crossing LOOP. */
348 /* Check for any intervening guards (includes conflicting loads). */
349 for (ir
= IR(J
->cur
.nins
-1); ir
> store
; ir
--)
350 if (irt_isguard(ir
->t
))
351 goto doemit
; /* No elimination possible. */
352 /* Remove redundant store from chain and replace with NOP. */
355 store
->t
.irt
= IRT_NIL
;
356 store
->op1
= store
->op2
= 0;
358 /* Now emit the new store instead. */
362 ref
= *(refp
= &store
->prev
);
365 return EMITFOLD
; /* Otherwise we have a conflict or simply no match. */
368 /* -- ULOAD forwarding ---------------------------------------------------- */
370 /* The current alias analysis for upvalues is very simplistic. It only
371 ** disambiguates between the unique upvalues of the same function.
372 ** This is good enough for now, since most upvalues are read-only.
374 ** A more precise analysis would be feasible with the help of the parser:
375 ** generate a unique key for every upvalue, even across all prototypes.
376 ** Lacking a realistic use-case, it's unclear whether this is beneficial.
378 static AliasRet
aa_uref(IRIns
*refa
, IRIns
*refb
)
380 if (refa
->o
!= refb
->o
)
381 return ALIAS_NO
; /* Different UREFx type. */
382 if (refa
->op1
== refb
->op1
) { /* Same function. */
383 if (refa
->op2
== refb
->op2
)
384 return ALIAS_MUST
; /* Same function, same upvalue idx. */
386 return ALIAS_NO
; /* Same function, different upvalue idx. */
387 } else { /* Different functions, check disambiguation hash values. */
388 if (((refa
->op2
^ refb
->op2
) & 0xff))
389 return ALIAS_NO
; /* Upvalues with different hash values cannot alias. */
391 return ALIAS_MAY
; /* No conclusion can be drawn for same hash value. */
395 /* ULOAD forwarding. */
396 TRef LJ_FASTCALL
lj_opt_fwd_uload(jit_State
*J
)
398 IRRef uref
= fins
->op1
;
399 IRRef lim
= uref
; /* Search limit. */
400 IRIns
*xr
= IR(uref
);
403 /* Search for conflicting stores. */
404 ref
= J
->chain
[IR_USTORE
];
406 IRIns
*store
= IR(ref
);
407 switch (aa_uref(xr
, IR(store
->op1
))) {
408 case ALIAS_NO
: break; /* Continue searching. */
409 case ALIAS_MAY
: lim
= ref
; goto cselim
; /* Limit search for load. */
410 case ALIAS_MUST
: return store
->op2
; /* Store forwarding. */
416 /* Try to find a matching load. Below the conflicting store, if any. */
417 return lj_opt_cselim(J
, lim
);
420 /* USTORE elimination. */
421 TRef LJ_FASTCALL
lj_opt_dse_ustore(jit_State
*J
)
423 IRRef xref
= fins
->op1
; /* xREF reference. */
424 IRRef val
= fins
->op2
; /* Stored value reference. */
425 IRIns
*xr
= IR(xref
);
426 IRRef1
*refp
= &J
->chain
[IR_USTORE
];
428 while (ref
> xref
) { /* Search for redundant or conflicting stores. */
429 IRIns
*store
= IR(ref
);
430 switch (aa_uref(xr
, IR(store
->op1
))) {
432 break; /* Continue searching. */
433 case ALIAS_MAY
: /* Store to MAYBE the same location. */
434 if (store
->op2
!= val
) /* Conflict if the value is different. */
436 break; /* Otherwise continue searching. */
437 case ALIAS_MUST
: /* Store to the same location. */
438 if (store
->op2
== val
) /* Same value: drop the new store. */
440 /* Different value: try to eliminate the redundant store. */
441 if (ref
> J
->chain
[IR_LOOP
]) { /* Quick check to avoid crossing LOOP. */
443 /* Check for any intervening guards (includes conflicting loads). */
444 for (ir
= IR(J
->cur
.nins
-1); ir
> store
; ir
--)
445 if (irt_isguard(ir
->t
))
446 goto doemit
; /* No elimination possible. */
447 /* Remove redundant store from chain and replace with NOP. */
450 store
->t
.irt
= IRT_NIL
;
451 store
->op1
= store
->op2
= 0;
453 if (ref
+1 < J
->cur
.nins
&&
454 store
[1].o
== IR_OBAR
&& store
[1].op1
== xref
) {
455 IRRef1
*bp
= &J
->chain
[IR_OBAR
];
457 for (obar
= IR(*bp
); *bp
> ref
+1; obar
= IR(*bp
))
459 /* Remove OBAR, too. */
462 obar
->t
.irt
= IRT_NIL
;
463 obar
->op1
= obar
->op2
= 0;
466 /* Now emit the new store instead. */
470 ref
= *(refp
= &store
->prev
);
473 return EMITFOLD
; /* Otherwise we have a conflict or simply no match. */
476 /* -- FLOAD forwarding and FSTORE elimination ----------------------------- */
478 /* Alias analysis for field access.
479 ** Field loads are cheap and field stores are rare.
480 ** Simple disambiguation based on field types is good enough.
482 static AliasRet
aa_fref(jit_State
*J
, IRIns
*refa
, IRIns
*refb
)
484 if (refa
->op2
!= refb
->op2
)
485 return ALIAS_NO
; /* Different fields. */
486 if (refa
->op1
== refb
->op1
)
487 return ALIAS_MUST
; /* Same field, same object. */
488 else if (refa
->op2
>= IRFL_TAB_META
&& refa
->op2
<= IRFL_TAB_NOMM
)
489 return aa_table(J
, refa
->op1
, refb
->op1
); /* Disambiguate tables. */
491 return ALIAS_MAY
; /* Same field, possibly different object. */
494 /* Only the loads for mutable fields end up here (see FOLD). */
495 TRef LJ_FASTCALL
lj_opt_fwd_fload(jit_State
*J
)
497 IRRef oref
= fins
->op1
; /* Object reference. */
498 IRRef fid
= fins
->op2
; /* Field ID. */
499 IRRef lim
= oref
; /* Search limit. */
502 /* Search for conflicting stores. */
503 ref
= J
->chain
[IR_FSTORE
];
505 IRIns
*store
= IR(ref
);
506 switch (aa_fref(J
, fins
, IR(store
->op1
))) {
507 case ALIAS_NO
: break; /* Continue searching. */
508 case ALIAS_MAY
: lim
= ref
; goto cselim
; /* Limit search for load. */
509 case ALIAS_MUST
: return store
->op2
; /* Store forwarding. */
514 /* No conflicting store: const-fold field loads from allocations. */
515 if (fid
== IRFL_TAB_META
) {
516 IRIns
*ir
= IR(oref
);
517 if (ir
->o
== IR_TNEW
|| ir
->o
== IR_TDUP
)
518 return lj_ir_knull(J
, IRT_TAB
);
522 /* Try to find a matching load. Below the conflicting store, if any. */
523 return lj_opt_cselim(J
, lim
);
526 /* FSTORE elimination. */
527 TRef LJ_FASTCALL
lj_opt_dse_fstore(jit_State
*J
)
529 IRRef fref
= fins
->op1
; /* FREF reference. */
530 IRRef val
= fins
->op2
; /* Stored value reference. */
531 IRIns
*xr
= IR(fref
);
532 IRRef1
*refp
= &J
->chain
[IR_FSTORE
];
534 while (ref
> fref
) { /* Search for redundant or conflicting stores. */
535 IRIns
*store
= IR(ref
);
536 switch (aa_fref(J
, xr
, IR(store
->op1
))) {
538 break; /* Continue searching. */
540 if (store
->op2
!= val
) /* Conflict if the value is different. */
542 break; /* Otherwise continue searching. */
544 if (store
->op2
== val
) /* Same value: drop the new store. */
546 /* Different value: try to eliminate the redundant store. */
547 if (ref
> J
->chain
[IR_LOOP
]) { /* Quick check to avoid crossing LOOP. */
549 /* Check for any intervening guards or conflicting loads. */
550 for (ir
= IR(J
->cur
.nins
-1); ir
> store
; ir
--)
551 if (irt_isguard(ir
->t
) || (ir
->o
== IR_FLOAD
&& ir
->op2
== xr
->op2
))
552 goto doemit
; /* No elimination possible. */
553 /* Remove redundant store from chain and replace with NOP. */
556 store
->t
.irt
= IRT_NIL
;
557 store
->op1
= store
->op2
= 0;
559 /* Now emit the new store instead. */
563 ref
= *(refp
= &store
->prev
);
566 return EMITFOLD
; /* Otherwise we have a conflict or simply no match. */
569 /* -- XLOAD forwarding and XSTORE elimination ----------------------------- */
571 /* Find cdata allocation for a reference (if any). */
572 static IRIns
*aa_findcnew(jit_State
*J
, IRIns
*ir
)
574 while (ir
->o
== IR_ADD
) {
575 if (!irref_isk(ir
->op1
)) {
576 IRIns
*ir1
= aa_findcnew(J
, IR(ir
->op1
)); /* Left-recursion. */
579 if (irref_isk(ir
->op2
)) return NULL
;
580 ir
= IR(ir
->op2
); /* Flatten right-recursion. */
582 return ir
->o
== IR_CNEW
? ir
: NULL
;
585 /* Alias analysis for two cdata allocations. */
586 static AliasRet
aa_cnew(jit_State
*J
, IRIns
*refa
, IRIns
*refb
)
588 IRIns
*cnewa
= aa_findcnew(J
, refa
);
589 IRIns
*cnewb
= aa_findcnew(J
, refb
);
591 return ALIAS_MAY
; /* Same allocation or neither is an allocation. */
593 return ALIAS_NO
; /* Two different allocations never alias. */
594 if (cnewb
) { cnewa
= cnewb
; refb
= refa
; }
595 return aa_escape(J
, cnewa
, refb
);
598 /* Alias analysis for XLOAD/XSTORE. */
599 static AliasRet
aa_xref(jit_State
*J
, IRIns
*refa
, IRIns
*xa
, IRIns
*xb
)
601 ptrdiff_t ofsa
= 0, ofsb
= 0;
602 IRIns
*refb
= IR(xb
->op1
);
603 IRIns
*basea
= refa
, *baseb
= refb
;
604 if (refa
== refb
&& irt_sametype(xa
->t
, xb
->t
))
605 return ALIAS_MUST
; /* Shortcut for same refs with identical type. */
606 /* Offset-based disambiguation. */
607 if (refa
->o
== IR_ADD
&& irref_isk(refa
->op2
)) {
608 IRIns
*irk
= IR(refa
->op2
);
609 basea
= IR(refa
->op1
);
610 ofsa
= (LJ_64
&& irk
->o
== IR_KINT64
) ? (ptrdiff_t)ir_k64(irk
)->u64
:
612 if (basea
== refb
&& ofsa
!= 0)
613 return ALIAS_NO
; /* base+-ofs vs. base. */
615 if (refb
->o
== IR_ADD
&& irref_isk(refb
->op2
)) {
616 IRIns
*irk
= IR(refb
->op2
);
617 baseb
= IR(refb
->op1
);
618 ofsb
= (LJ_64
&& irk
->o
== IR_KINT64
) ? (ptrdiff_t)ir_k64(irk
)->u64
:
620 if (refa
== baseb
&& ofsb
!= 0)
621 return ALIAS_NO
; /* base vs. base+-ofs. */
623 /* This implements (very) strict aliasing rules.
624 ** Different types do NOT alias, except for differences in signedness.
625 ** Type punning through unions is allowed (but forces a reload).
627 if (basea
== baseb
) {
628 ptrdiff_t sza
= irt_size(xa
->t
), szb
= irt_size(xb
->t
);
630 if (sza
== szb
&& irt_isfp(xa
->t
) == irt_isfp(xb
->t
))
631 return ALIAS_MUST
; /* Same-sized, same-kind. May need to convert. */
632 } else if (ofsa
+ sza
<= ofsb
|| ofsb
+ szb
<= ofsa
) {
633 return ALIAS_NO
; /* Non-overlapping base+-o1 vs. base+-o2. */
635 /* NYI: extract, extend or reinterpret bits (int <-> fp). */
636 return ALIAS_MAY
; /* Overlapping or type punning: force reload. */
638 if (!irt_sametype(xa
->t
, xb
->t
) &&
639 !(irt_typerange(xa
->t
, IRT_I8
, IRT_U64
) &&
640 ((xa
->t
.irt
- IRT_I8
) ^ (xb
->t
.irt
- IRT_I8
)) == 1))
642 /* NYI: structural disambiguation. */
643 return aa_cnew(J
, basea
, baseb
); /* Try to disambiguate allocations. */
646 /* Return CSEd reference or 0. Caveat: swaps lower ref to the right! */
647 static IRRef
reassoc_trycse(jit_State
*J
, IROp op
, IRRef op1
, IRRef op2
)
649 IRRef ref
= J
->chain
[op
];
651 if (op2
> lim
) { lim
= op2
; op2
= op1
; op1
= lim
; }
654 if (ir
->op1
== op1
&& ir
->op2
== op2
)
661 /* Reassociate index references. */
662 static IRRef
reassoc_xref(jit_State
*J
, IRIns
*ir
)
665 if (ir
->o
== IR_ADD
&& irref_isk(ir
->op2
)) { /* Get constant offset. */
666 IRIns
*irk
= IR(ir
->op2
);
667 ofs
= (LJ_64
&& irk
->o
== IR_KINT64
) ? (ptrdiff_t)ir_k64(irk
)->u64
:
671 if (ir
->o
== IR_ADD
) { /* Add of base + index. */
672 /* Index ref > base ref for loop-carried dependences. Only check op1. */
673 IRIns
*ir2
, *ir1
= IR(ir
->op1
);
676 /* Determine index shifts. Don't bother with IR_MUL here. */
677 if (ir1
->o
== IR_BSHL
&& irref_isk(ir1
->op2
))
678 shift
= IR(ir1
->op2
)->i
;
679 else if (ir1
->o
== IR_ADD
&& ir1
->op1
== ir1
->op2
)
684 /* A non-reassociated add. Must be a loop-carried dependence. */
685 if (ir2
->o
== IR_ADD
&& irt_isint(ir2
->t
) && irref_isk(ir2
->op2
))
686 ofs
+= (ptrdiff_t)IR(ir2
->op2
)->i
<< shift
;
690 /* Try to CSE the reassociated chain. Give up if not found. */
692 !(idxref
= reassoc_trycse(J
, ir1
->o
, idxref
,
693 ir1
->o
== IR_BSHL
? ir1
->op2
: idxref
)))
695 if (!(idxref
= reassoc_trycse(J
, IR_ADD
, idxref
, ir
->op2
)))
698 IRRef refk
= tref_ref(lj_ir_kintp(J
, ofs
));
699 if (!(idxref
= reassoc_trycse(J
, IR_ADD
, idxref
, refk
)))
702 return idxref
; /* Success, found a reassociated index reference. Phew. */
704 return 0; /* Failure. */
707 /* XLOAD forwarding. */
708 TRef LJ_FASTCALL
lj_opt_fwd_xload(jit_State
*J
)
710 IRRef xref
= fins
->op1
;
711 IRIns
*xr
= IR(xref
);
712 IRRef lim
= xref
; /* Search limit. */
715 if ((fins
->op2
& IRXLOAD_READONLY
))
717 if ((fins
->op2
& IRXLOAD_VOLATILE
))
720 /* Search for conflicting stores. */
721 ref
= J
->chain
[IR_XSTORE
];
723 if (J
->chain
[IR_CALLXS
] > lim
) lim
= J
->chain
[IR_CALLXS
];
724 if (J
->chain
[IR_XBAR
] > lim
) lim
= J
->chain
[IR_XBAR
];
726 IRIns
*store
= IR(ref
);
727 switch (aa_xref(J
, xr
, fins
, store
)) {
728 case ALIAS_NO
: break; /* Continue searching. */
729 case ALIAS_MAY
: lim
= ref
; goto cselim
; /* Limit search for load. */
731 /* Emit conversion if the loaded type doesn't match the forwarded type. */
732 if (!irt_sametype(fins
->t
, IR(store
->op2
)->t
)) {
733 IRType st
= irt_type(fins
->t
);
734 if (st
== IRT_I8
|| st
== IRT_I16
) { /* Trunc + sign-extend. */
736 } else if (st
== IRT_U8
|| st
== IRT_U16
) { /* Trunc + zero-extend. */
737 } else if (st
== IRT_INT
) {
738 st
= irt_type(IR(store
->op2
)->t
); /* Needs dummy CONV.int.*. */
739 } else { /* I64/U64 are boxed, U32 is hidden behind a CONV.num.u32. */
742 fins
->ot
= IRTI(IR_CONV
);
743 fins
->op1
= store
->op2
;
744 fins
->op2
= (IRT_INT
<<5)|st
;
748 return store
->op2
; /* Store forwarding. */
754 /* Try to find a matching load. Below the conflicting store, if any. */
755 ref
= J
->chain
[IR_XLOAD
];
757 /* CSE for XLOAD depends on the type, but not on the IRXLOAD_* flags. */
758 if (IR(ref
)->op1
== xref
&& irt_sametype(IR(ref
)->t
, fins
->t
))
763 /* Reassociate XLOAD across PHIs to handle a[i-1] forwarding case. */
764 if (!(fins
->op2
& IRXLOAD_READONLY
) && J
->chain
[IR_LOOP
] &&
765 xref
== fins
->op1
&& (xref
= reassoc_xref(J
, xr
)) != 0) {
766 ref
= J
->chain
[IR_XSTORE
];
767 while (ref
> lim
) /* Skip stores that have already been checked. */
771 goto retry
; /* Retry with the reassociated reference. */
777 /* XSTORE elimination. */
778 TRef LJ_FASTCALL
lj_opt_dse_xstore(jit_State
*J
)
780 IRRef xref
= fins
->op1
;
781 IRIns
*xr
= IR(xref
);
782 IRRef lim
= xref
; /* Search limit. */
783 IRRef val
= fins
->op2
; /* Stored value reference. */
784 IRRef1
*refp
= &J
->chain
[IR_XSTORE
];
786 if (J
->chain
[IR_CALLXS
] > lim
) lim
= J
->chain
[IR_CALLXS
];
787 if (J
->chain
[IR_XBAR
] > lim
) lim
= J
->chain
[IR_XBAR
];
788 while (ref
> lim
) { /* Search for redundant or conflicting stores. */
789 IRIns
*store
= IR(ref
);
790 switch (aa_xref(J
, xr
, fins
, store
)) {
792 break; /* Continue searching. */
794 if (store
->op2
!= val
) /* Conflict if the value is different. */
796 break; /* Otherwise continue searching. */
798 if (store
->op2
== val
) /* Same value: drop the new store. */
800 /* Different value: try to eliminate the redundant store. */
801 if (ref
> J
->chain
[IR_LOOP
]) { /* Quick check to avoid crossing LOOP. */
803 /* Check for any intervening guards or any XLOADs (no AA performed). */
804 for (ir
= IR(J
->cur
.nins
-1); ir
> store
; ir
--)
805 if (irt_isguard(ir
->t
) || ir
->o
== IR_XLOAD
)
806 goto doemit
; /* No elimination possible. */
807 /* Remove redundant store from chain and replace with NOP. */
810 store
->t
.irt
= IRT_NIL
;
811 store
->op1
= store
->op2
= 0;
813 /* Now emit the new store instead. */
817 ref
= *(refp
= &store
->prev
);
820 return EMITFOLD
; /* Otherwise we have a conflict or simply no match. */
823 /* -- Forwarding of lj_tab_len -------------------------------------------- */
825 /* This is rather simplistic right now, but better than nothing. */
826 TRef LJ_FASTCALL
lj_opt_fwd_tab_len(jit_State
*J
)
828 IRRef tab
= fins
->op1
; /* Table reference. */
829 IRRef lim
= tab
; /* Search limit. */
832 /* Any ASTORE is a conflict and limits the search. */
833 if (J
->chain
[IR_ASTORE
] > lim
) lim
= J
->chain
[IR_ASTORE
];
835 /* Search for conflicting HSTORE with numeric key. */
836 ref
= J
->chain
[IR_HSTORE
];
838 IRIns
*store
= IR(ref
);
839 IRIns
*href
= IR(store
->op1
);
840 IRIns
*key
= IR(href
->op2
);
841 if (irt_isnum(key
->o
== IR_KSLOT
? IR(key
->op1
)->t
: key
->t
)) {
842 lim
= ref
; /* Conflicting store found, limits search for TLEN. */
848 /* Try to find a matching load. Below the conflicting store, if any. */
849 return lj_opt_cselim(J
, lim
);
852 /* -- ASTORE/HSTORE previous type analysis -------------------------------- */
854 /* Check whether the previous value for a table store is non-nil.
855 ** This can be derived either from a previous store or from a previous
856 ** load (because all loads from tables perform a type check).
858 ** The result of the analysis can be used to avoid the metatable check
859 ** and the guard against HREF returning niltv. Both of these are cheap,
860 ** so let's not spend too much effort on the analysis.
862 ** A result of 1 is exact: previous value CANNOT be nil.
863 ** A result of 0 is inexact: previous value MAY be nil.
865 int lj_opt_fwd_wasnonnil(jit_State
*J
, IROpT loadop
, IRRef xref
)
867 /* First check stores. */
868 IRRef ref
= J
->chain
[loadop
+IRDELTA_L2S
];
870 IRIns
*store
= IR(ref
);
871 if (store
->op1
== xref
) { /* Same xREF. */
872 /* A nil store MAY alias, but a non-nil store MUST alias. */
873 return !irt_isnil(store
->t
);
874 } else if (irt_isnil(store
->t
)) { /* Must check any nil store. */
875 IRRef skref
= IR(store
->op1
)->op2
;
876 IRRef xkref
= IR(xref
)->op2
;
877 /* Same key type MAY alias. Need ALOAD check due to multiple int types. */
878 if (loadop
== IR_ALOAD
|| irt_sametype(IR(skref
)->t
, IR(xkref
)->t
)) {
879 if (skref
== xkref
|| !irref_isk(skref
) || !irref_isk(xkref
))
880 return 0; /* A nil store with same const key or var key MAY alias. */
881 /* Different const keys CANNOT alias. */
882 } /* Different key types CANNOT alias. */
883 } /* Other non-nil stores MAY alias. */
887 /* Check loads since nothing could be derived from stores. */
888 ref
= J
->chain
[loadop
];
890 IRIns
*load
= IR(ref
);
891 if (load
->op1
== xref
) { /* Same xREF. */
892 /* A nil load MAY alias, but a non-nil load MUST alias. */
893 return !irt_isnil(load
->t
);
894 } /* Other non-nil loads MAY alias. */
897 return 0; /* Nothing derived at all, previous value MAY be nil. */
900 /* ------------------------------------------------------------------------ */