2 * Optimizations for Tiny Code Generator for QEMU
4 * Copyright (c) 2010 Samsung Electronics.
5 * Contributed by Kirill Batuzov <batuzovk@ispras.ru>
7 * Permission is hereby granted, free of charge, to any person obtaining a copy
8 * of this software and associated documentation files (the "Software"), to deal
9 * in the Software without restriction, including without limitation the rights
10 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
11 * copies of the Software, and to permit persons to whom the Software is
12 * furnished to do so, subject to the following conditions:
14 * The above copyright notice and this permission notice shall be included in
15 * all copies or substantial portions of the Software.
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
31 #include "qemu-common.h"
34 #define CASE_OP_32_64(x) \
35 glue(glue(case INDEX_op_, x), _i32): \
36 glue(glue(case INDEX_op_, x), _i64)
44 struct tcg_temp_info
{
49 tcg_target_ulong mask
;
52 static struct tcg_temp_info temps
[TCG_MAX_TEMPS
];
54 /* Reset TEMP's state to TCG_TEMP_UNDEF. If TEMP only had one copy, remove
55 the copy flag from the left temp. */
56 static void reset_temp(TCGArg temp
)
58 if (temps
[temp
].state
== TCG_TEMP_COPY
) {
59 if (temps
[temp
].prev_copy
== temps
[temp
].next_copy
) {
60 temps
[temps
[temp
].next_copy
].state
= TCG_TEMP_UNDEF
;
62 temps
[temps
[temp
].next_copy
].prev_copy
= temps
[temp
].prev_copy
;
63 temps
[temps
[temp
].prev_copy
].next_copy
= temps
[temp
].next_copy
;
66 temps
[temp
].state
= TCG_TEMP_UNDEF
;
67 temps
[temp
].mask
= -1;
70 /* Reset all temporaries, given that there are NB_TEMPS of them. */
71 static void reset_all_temps(int nb_temps
)
74 for (i
= 0; i
< nb_temps
; i
++) {
75 temps
[i
].state
= TCG_TEMP_UNDEF
;
80 static int op_bits(TCGOpcode op
)
82 const TCGOpDef
*def
= &tcg_op_defs
[op
];
83 return def
->flags
& TCG_OPF_64BIT
? 64 : 32;
86 static TCGOpcode
op_to_movi(TCGOpcode op
)
88 switch (op_bits(op
)) {
90 return INDEX_op_movi_i32
;
92 return INDEX_op_movi_i64
;
94 fprintf(stderr
, "op_to_movi: unexpected return value of "
95 "function op_bits.\n");
100 static TCGArg
find_better_copy(TCGContext
*s
, TCGArg temp
)
104 /* If this is already a global, we can't do better. */
105 if (temp
< s
->nb_globals
) {
109 /* Search for a global first. */
110 for (i
= temps
[temp
].next_copy
; i
!= temp
; i
= temps
[i
].next_copy
) {
111 if (i
< s
->nb_globals
) {
116 /* If it is a temp, search for a temp local. */
117 if (!s
->temps
[temp
].temp_local
) {
118 for (i
= temps
[temp
].next_copy
; i
!= temp
; i
= temps
[i
].next_copy
) {
119 if (s
->temps
[i
].temp_local
) {
125 /* Failure to find a better representation, return the same temp. */
129 static bool temps_are_copies(TCGArg arg1
, TCGArg arg2
)
137 if (temps
[arg1
].state
!= TCG_TEMP_COPY
138 || temps
[arg2
].state
!= TCG_TEMP_COPY
) {
142 for (i
= temps
[arg1
].next_copy
; i
!= arg1
; i
= temps
[i
].next_copy
) {
151 static void tcg_opt_gen_mov(TCGContext
*s
, TCGArg
*gen_args
,
152 TCGArg dst
, TCGArg src
)
155 temps
[dst
].mask
= temps
[src
].mask
;
156 assert(temps
[src
].state
!= TCG_TEMP_CONST
);
158 if (s
->temps
[src
].type
== s
->temps
[dst
].type
) {
159 if (temps
[src
].state
!= TCG_TEMP_COPY
) {
160 temps
[src
].state
= TCG_TEMP_COPY
;
161 temps
[src
].next_copy
= src
;
162 temps
[src
].prev_copy
= src
;
164 temps
[dst
].state
= TCG_TEMP_COPY
;
165 temps
[dst
].next_copy
= temps
[src
].next_copy
;
166 temps
[dst
].prev_copy
= src
;
167 temps
[temps
[dst
].next_copy
].prev_copy
= dst
;
168 temps
[src
].next_copy
= dst
;
175 static void tcg_opt_gen_movi(TCGArg
*gen_args
, TCGArg dst
, TCGArg val
)
178 temps
[dst
].state
= TCG_TEMP_CONST
;
179 temps
[dst
].val
= val
;
180 temps
[dst
].mask
= val
;
185 static TCGOpcode
op_to_mov(TCGOpcode op
)
187 switch (op_bits(op
)) {
189 return INDEX_op_mov_i32
;
191 return INDEX_op_mov_i64
;
193 fprintf(stderr
, "op_to_mov: unexpected return value of "
194 "function op_bits.\n");
199 static TCGArg
do_constant_folding_2(TCGOpcode op
, TCGArg x
, TCGArg y
)
222 case INDEX_op_shl_i32
:
223 return (uint32_t)x
<< (uint32_t)y
;
225 case INDEX_op_shl_i64
:
226 return (uint64_t)x
<< (uint64_t)y
;
228 case INDEX_op_shr_i32
:
229 return (uint32_t)x
>> (uint32_t)y
;
231 case INDEX_op_shr_i64
:
232 return (uint64_t)x
>> (uint64_t)y
;
234 case INDEX_op_sar_i32
:
235 return (int32_t)x
>> (int32_t)y
;
237 case INDEX_op_sar_i64
:
238 return (int64_t)x
>> (int64_t)y
;
240 case INDEX_op_rotr_i32
:
243 case INDEX_op_rotr_i64
:
246 case INDEX_op_rotl_i32
:
249 case INDEX_op_rotl_i64
:
273 CASE_OP_32_64(ext8s
):
276 CASE_OP_32_64(ext16s
):
279 CASE_OP_32_64(ext8u
):
282 CASE_OP_32_64(ext16u
):
285 case INDEX_op_ext32s_i64
:
288 case INDEX_op_ext32u_i64
:
291 case INDEX_op_muluh_i32
:
292 return ((uint64_t)(uint32_t)x
* (uint32_t)y
) >> 32;
293 case INDEX_op_mulsh_i32
:
294 return ((int64_t)(int32_t)x
* (int32_t)y
) >> 32;
296 case INDEX_op_muluh_i64
:
297 mulu64(&l64
, &h64
, x
, y
);
299 case INDEX_op_mulsh_i64
:
300 muls64(&l64
, &h64
, x
, y
);
303 case INDEX_op_div_i32
:
304 /* Avoid crashing on divide by zero, otherwise undefined. */
305 return (int32_t)x
/ ((int32_t)y
? : 1);
306 case INDEX_op_divu_i32
:
307 return (uint32_t)x
/ ((uint32_t)y
? : 1);
308 case INDEX_op_div_i64
:
309 return (int64_t)x
/ ((int64_t)y
? : 1);
310 case INDEX_op_divu_i64
:
311 return (uint64_t)x
/ ((uint64_t)y
? : 1);
313 case INDEX_op_rem_i32
:
314 return (int32_t)x
% ((int32_t)y
? : 1);
315 case INDEX_op_remu_i32
:
316 return (uint32_t)x
% ((uint32_t)y
? : 1);
317 case INDEX_op_rem_i64
:
318 return (int64_t)x
% ((int64_t)y
? : 1);
319 case INDEX_op_remu_i64
:
320 return (uint64_t)x
% ((uint64_t)y
? : 1);
324 "Unrecognized operation %d in do_constant_folding.\n", op
);
329 static TCGArg
do_constant_folding(TCGOpcode op
, TCGArg x
, TCGArg y
)
331 TCGArg res
= do_constant_folding_2(op
, x
, y
);
332 if (op_bits(op
) == 32) {
338 static bool do_constant_folding_cond_32(uint32_t x
, uint32_t y
, TCGCond c
)
346 return (int32_t)x
< (int32_t)y
;
348 return (int32_t)x
>= (int32_t)y
;
350 return (int32_t)x
<= (int32_t)y
;
352 return (int32_t)x
> (int32_t)y
;
366 static bool do_constant_folding_cond_64(uint64_t x
, uint64_t y
, TCGCond c
)
374 return (int64_t)x
< (int64_t)y
;
376 return (int64_t)x
>= (int64_t)y
;
378 return (int64_t)x
<= (int64_t)y
;
380 return (int64_t)x
> (int64_t)y
;
394 static bool do_constant_folding_cond_eq(TCGCond c
)
414 /* Return 2 if the condition can't be simplified, and the result
415 of the condition (0 or 1) if it can */
416 static TCGArg
do_constant_folding_cond(TCGOpcode op
, TCGArg x
,
419 if (temps
[x
].state
== TCG_TEMP_CONST
&& temps
[y
].state
== TCG_TEMP_CONST
) {
420 switch (op_bits(op
)) {
422 return do_constant_folding_cond_32(temps
[x
].val
, temps
[y
].val
, c
);
424 return do_constant_folding_cond_64(temps
[x
].val
, temps
[y
].val
, c
);
428 } else if (temps_are_copies(x
, y
)) {
429 return do_constant_folding_cond_eq(c
);
430 } else if (temps
[y
].state
== TCG_TEMP_CONST
&& temps
[y
].val
== 0) {
444 /* Return 2 if the condition can't be simplified, and the result
445 of the condition (0 or 1) if it can */
446 static TCGArg
do_constant_folding_cond2(TCGArg
*p1
, TCGArg
*p2
, TCGCond c
)
448 TCGArg al
= p1
[0], ah
= p1
[1];
449 TCGArg bl
= p2
[0], bh
= p2
[1];
451 if (temps
[bl
].state
== TCG_TEMP_CONST
452 && temps
[bh
].state
== TCG_TEMP_CONST
) {
453 uint64_t b
= ((uint64_t)temps
[bh
].val
<< 32) | (uint32_t)temps
[bl
].val
;
455 if (temps
[al
].state
== TCG_TEMP_CONST
456 && temps
[ah
].state
== TCG_TEMP_CONST
) {
458 a
= ((uint64_t)temps
[ah
].val
<< 32) | (uint32_t)temps
[al
].val
;
459 return do_constant_folding_cond_64(a
, b
, c
);
472 if (temps_are_copies(al
, bl
) && temps_are_copies(ah
, bh
)) {
473 return do_constant_folding_cond_eq(c
);
478 static bool swap_commutative(TCGArg dest
, TCGArg
*p1
, TCGArg
*p2
)
480 TCGArg a1
= *p1
, a2
= *p2
;
482 sum
+= temps
[a1
].state
== TCG_TEMP_CONST
;
483 sum
-= temps
[a2
].state
== TCG_TEMP_CONST
;
485 /* Prefer the constant in second argument, and then the form
486 op a, a, b, which is better handled on non-RISC hosts. */
487 if (sum
> 0 || (sum
== 0 && dest
== a2
)) {
495 static bool swap_commutative2(TCGArg
*p1
, TCGArg
*p2
)
498 sum
+= temps
[p1
[0]].state
== TCG_TEMP_CONST
;
499 sum
+= temps
[p1
[1]].state
== TCG_TEMP_CONST
;
500 sum
-= temps
[p2
[0]].state
== TCG_TEMP_CONST
;
501 sum
-= temps
[p2
[1]].state
== TCG_TEMP_CONST
;
504 t
= p1
[0], p1
[0] = p2
[0], p2
[0] = t
;
505 t
= p1
[1], p1
[1] = p2
[1], p2
[1] = t
;
511 /* Propagate constants and copies, fold constant expressions. */
512 static TCGArg
*tcg_constant_folding(TCGContext
*s
, uint16_t *tcg_opc_ptr
,
513 TCGArg
*args
, TCGOpDef
*tcg_op_defs
)
515 int i
, nb_ops
, op_index
, nb_temps
, nb_globals
, nb_call_args
;
516 tcg_target_ulong mask
, affected
;
522 /* Array VALS has an element for each temp.
523 If this temp holds a constant then its value is kept in VALS' element.
524 If this temp is a copy of other ones then the other copies are
525 available through the doubly linked circular list. */
527 nb_temps
= s
->nb_temps
;
528 nb_globals
= s
->nb_globals
;
529 reset_all_temps(nb_temps
);
531 nb_ops
= tcg_opc_ptr
- s
->gen_opc_buf
;
533 for (op_index
= 0; op_index
< nb_ops
; op_index
++) {
534 op
= s
->gen_opc_buf
[op_index
];
535 def
= &tcg_op_defs
[op
];
536 /* Do copy propagation */
537 if (op
== INDEX_op_call
) {
538 int nb_oargs
= args
[0] >> 16;
539 int nb_iargs
= args
[0] & 0xffff;
540 for (i
= nb_oargs
+ 1; i
< nb_oargs
+ nb_iargs
+ 1; i
++) {
541 if (temps
[args
[i
]].state
== TCG_TEMP_COPY
) {
542 args
[i
] = find_better_copy(s
, args
[i
]);
546 for (i
= def
->nb_oargs
; i
< def
->nb_oargs
+ def
->nb_iargs
; i
++) {
547 if (temps
[args
[i
]].state
== TCG_TEMP_COPY
) {
548 args
[i
] = find_better_copy(s
, args
[i
]);
553 /* For commutative operations make constant second argument */
563 CASE_OP_32_64(muluh
):
564 CASE_OP_32_64(mulsh
):
565 swap_commutative(args
[0], &args
[1], &args
[2]);
567 CASE_OP_32_64(brcond
):
568 if (swap_commutative(-1, &args
[0], &args
[1])) {
569 args
[2] = tcg_swap_cond(args
[2]);
572 CASE_OP_32_64(setcond
):
573 if (swap_commutative(args
[0], &args
[1], &args
[2])) {
574 args
[3] = tcg_swap_cond(args
[3]);
577 CASE_OP_32_64(movcond
):
578 if (swap_commutative(-1, &args
[1], &args
[2])) {
579 args
[5] = tcg_swap_cond(args
[5]);
581 /* For movcond, we canonicalize the "false" input reg to match
582 the destination reg so that the tcg backend can implement
583 a "move if true" operation. */
584 if (swap_commutative(args
[0], &args
[4], &args
[3])) {
585 args
[5] = tcg_invert_cond(args
[5]);
589 swap_commutative(args
[0], &args
[2], &args
[4]);
590 swap_commutative(args
[1], &args
[3], &args
[5]);
592 CASE_OP_32_64(mulu2
):
593 CASE_OP_32_64(muls2
):
594 swap_commutative(args
[0], &args
[2], &args
[3]);
596 case INDEX_op_brcond2_i32
:
597 if (swap_commutative2(&args
[0], &args
[2])) {
598 args
[4] = tcg_swap_cond(args
[4]);
601 case INDEX_op_setcond2_i32
:
602 if (swap_commutative2(&args
[1], &args
[3])) {
603 args
[5] = tcg_swap_cond(args
[5]);
610 /* Simplify expressions for "shift/rot r, 0, a => movi r, 0",
611 and "sub r, 0, a => neg r, a" case. */
618 if (temps
[args
[1]].state
== TCG_TEMP_CONST
619 && temps
[args
[1]].val
== 0) {
620 s
->gen_opc_buf
[op_index
] = op_to_movi(op
);
621 tcg_opt_gen_movi(gen_args
, args
[0], 0);
632 if (temps
[args
[2]].state
== TCG_TEMP_CONST
) {
633 /* Proceed with possible constant folding. */
636 if (op
== INDEX_op_sub_i32
) {
637 neg_op
= INDEX_op_neg_i32
;
638 have_neg
= TCG_TARGET_HAS_neg_i32
;
640 neg_op
= INDEX_op_neg_i64
;
641 have_neg
= TCG_TARGET_HAS_neg_i64
;
646 if (temps
[args
[1]].state
== TCG_TEMP_CONST
647 && temps
[args
[1]].val
== 0) {
648 s
->gen_opc_buf
[op_index
] = neg_op
;
650 gen_args
[0] = args
[0];
651 gen_args
[1] = args
[2];
660 if (temps
[args
[1]].state
!= TCG_TEMP_CONST
661 && temps
[args
[2]].state
== TCG_TEMP_CONST
662 && temps
[args
[2]].val
== -1) {
668 if (temps
[args
[1]].state
!= TCG_TEMP_CONST
669 && temps
[args
[2]].state
== TCG_TEMP_CONST
670 && temps
[args
[2]].val
== 0) {
676 if (temps
[args
[2]].state
!= TCG_TEMP_CONST
677 && temps
[args
[1]].state
== TCG_TEMP_CONST
678 && temps
[args
[1]].val
== -1) {
685 if (temps
[args
[2]].state
!= TCG_TEMP_CONST
686 && temps
[args
[1]].state
== TCG_TEMP_CONST
687 && temps
[args
[1]].val
== 0) {
697 if (def
->flags
& TCG_OPF_64BIT
) {
698 not_op
= INDEX_op_not_i64
;
699 have_not
= TCG_TARGET_HAS_not_i64
;
701 not_op
= INDEX_op_not_i32
;
702 have_not
= TCG_TARGET_HAS_not_i32
;
707 s
->gen_opc_buf
[op_index
] = not_op
;
709 gen_args
[0] = args
[0];
710 gen_args
[1] = args
[i
];
719 /* Simplify expression for "op r, a, const => mov r, a" cases */
731 if (temps
[args
[1]].state
!= TCG_TEMP_CONST
732 && temps
[args
[2]].state
== TCG_TEMP_CONST
733 && temps
[args
[2]].val
== 0) {
740 if (temps
[args
[1]].state
!= TCG_TEMP_CONST
741 && temps
[args
[2]].state
== TCG_TEMP_CONST
742 && temps
[args
[2]].val
== -1) {
747 if (temps_are_copies(args
[0], args
[1])) {
748 s
->gen_opc_buf
[op_index
] = INDEX_op_nop
;
750 s
->gen_opc_buf
[op_index
] = op_to_mov(op
);
751 tcg_opt_gen_mov(s
, gen_args
, args
[0], args
[1]);
760 /* Simplify using known-zero bits. Currently only ops with a single
761 output argument is supported. */
765 CASE_OP_32_64(ext8s
):
766 if ((temps
[args
[1]].mask
& 0x80) != 0) {
769 CASE_OP_32_64(ext8u
):
772 CASE_OP_32_64(ext16s
):
773 if ((temps
[args
[1]].mask
& 0x8000) != 0) {
776 CASE_OP_32_64(ext16u
):
779 case INDEX_op_ext32s_i64
:
780 if ((temps
[args
[1]].mask
& 0x80000000) != 0) {
783 case INDEX_op_ext32u_i64
:
788 mask
= temps
[args
[2]].mask
;
789 if (temps
[args
[2]].state
== TCG_TEMP_CONST
) {
791 affected
= temps
[args
[1]].mask
& ~mask
;
793 mask
= temps
[args
[1]].mask
& mask
;
797 /* Known-zeros does not imply known-ones. Therefore unless
798 args[2] is constant, we can't infer anything from it. */
799 if (temps
[args
[2]].state
== TCG_TEMP_CONST
) {
800 mask
= ~temps
[args
[2]].mask
;
803 /* But we certainly know nothing outside args[1] may be set. */
804 mask
= temps
[args
[1]].mask
;
807 case INDEX_op_sar_i32
:
808 if (temps
[args
[2]].state
== TCG_TEMP_CONST
) {
809 mask
= (int32_t)temps
[args
[1]].mask
>> temps
[args
[2]].val
;
812 case INDEX_op_sar_i64
:
813 if (temps
[args
[2]].state
== TCG_TEMP_CONST
) {
814 mask
= (int64_t)temps
[args
[1]].mask
>> temps
[args
[2]].val
;
818 case INDEX_op_shr_i32
:
819 if (temps
[args
[2]].state
== TCG_TEMP_CONST
) {
820 mask
= (uint32_t)temps
[args
[1]].mask
>> temps
[args
[2]].val
;
823 case INDEX_op_shr_i64
:
824 if (temps
[args
[2]].state
== TCG_TEMP_CONST
) {
825 mask
= (uint64_t)temps
[args
[1]].mask
>> temps
[args
[2]].val
;
830 if (temps
[args
[2]].state
== TCG_TEMP_CONST
) {
831 mask
= temps
[args
[1]].mask
<< temps
[args
[2]].val
;
836 /* Set to 1 all bits to the left of the rightmost. */
837 mask
= -(temps
[args
[1]].mask
& -temps
[args
[1]].mask
);
840 CASE_OP_32_64(deposit
):
841 tmp
= ((1ull << args
[4]) - 1);
842 mask
= ((temps
[args
[1]].mask
& ~(tmp
<< args
[3]))
843 | ((temps
[args
[2]].mask
& tmp
) << args
[3]));
848 mask
= temps
[args
[1]].mask
| temps
[args
[2]].mask
;
851 CASE_OP_32_64(setcond
):
855 CASE_OP_32_64(movcond
):
856 mask
= temps
[args
[3]].mask
| temps
[args
[4]].mask
;
860 case INDEX_op_qemu_ld8u
:
863 CASE_OP_32_64(ld16u
):
864 case INDEX_op_qemu_ld16u
:
867 case INDEX_op_ld32u_i64
:
868 #if TCG_TARGET_REG_BITS == 64
869 case INDEX_op_qemu_ld32u
:
874 CASE_OP_32_64(qemu_ld
):
876 TCGMemOp mop
= args
[def
->nb_oargs
+ def
->nb_iargs
];
877 if (!(mop
& MO_SIGN
)) {
878 mask
= (2ULL << ((8 << (mop
& MO_SIZE
)) - 1)) - 1;
887 /* 32-bit ops (non 64-bit ops and non load/store ops) generate 32-bit
889 if (!(def
->flags
& (TCG_OPF_CALL_CLOBBER
| TCG_OPF_64BIT
))) {
894 assert(def
->nb_oargs
== 1);
895 s
->gen_opc_buf
[op_index
] = op_to_movi(op
);
896 tcg_opt_gen_movi(gen_args
, args
[0], 0);
897 args
+= def
->nb_oargs
+ def
->nb_iargs
+ def
->nb_cargs
;
902 assert(def
->nb_oargs
== 1);
903 if (temps_are_copies(args
[0], args
[1])) {
904 s
->gen_opc_buf
[op_index
] = INDEX_op_nop
;
905 } else if (temps
[args
[1]].state
!= TCG_TEMP_CONST
) {
906 s
->gen_opc_buf
[op_index
] = op_to_mov(op
);
907 tcg_opt_gen_mov(s
, gen_args
, args
[0], args
[1]);
910 s
->gen_opc_buf
[op_index
] = op_to_movi(op
);
911 tcg_opt_gen_movi(gen_args
, args
[0], temps
[args
[1]].val
);
914 args
+= def
->nb_iargs
+ 1;
918 /* Simplify expression for "op r, a, 0 => movi r, 0" cases */
922 CASE_OP_32_64(muluh
):
923 CASE_OP_32_64(mulsh
):
924 if ((temps
[args
[2]].state
== TCG_TEMP_CONST
925 && temps
[args
[2]].val
== 0)) {
926 s
->gen_opc_buf
[op_index
] = op_to_movi(op
);
927 tcg_opt_gen_movi(gen_args
, args
[0], 0);
937 /* Simplify expression for "op r, a, a => mov r, a" cases */
941 if (temps_are_copies(args
[1], args
[2])) {
942 if (temps_are_copies(args
[0], args
[1])) {
943 s
->gen_opc_buf
[op_index
] = INDEX_op_nop
;
945 s
->gen_opc_buf
[op_index
] = op_to_mov(op
);
946 tcg_opt_gen_mov(s
, gen_args
, args
[0], args
[1]);
957 /* Simplify expression for "op r, a, a => movi r, 0" cases */
962 if (temps_are_copies(args
[1], args
[2])) {
963 s
->gen_opc_buf
[op_index
] = op_to_movi(op
);
964 tcg_opt_gen_movi(gen_args
, args
[0], 0);
974 /* Propagate constants through copy operations and do constant
975 folding. Constants will be substituted to arguments by register
976 allocator where needed and possible. Also detect copies. */
979 if (temps_are_copies(args
[0], args
[1])) {
981 s
->gen_opc_buf
[op_index
] = INDEX_op_nop
;
984 if (temps
[args
[1]].state
!= TCG_TEMP_CONST
) {
985 tcg_opt_gen_mov(s
, gen_args
, args
[0], args
[1]);
990 /* Source argument is constant. Rewrite the operation and
991 let movi case handle it. */
993 s
->gen_opc_buf
[op_index
] = op
;
994 args
[1] = temps
[args
[1]].val
;
997 tcg_opt_gen_movi(gen_args
, args
[0], args
[1]);
1004 CASE_OP_32_64(ext8s
):
1005 CASE_OP_32_64(ext8u
):
1006 CASE_OP_32_64(ext16s
):
1007 CASE_OP_32_64(ext16u
):
1008 case INDEX_op_ext32s_i64
:
1009 case INDEX_op_ext32u_i64
:
1010 if (temps
[args
[1]].state
== TCG_TEMP_CONST
) {
1011 s
->gen_opc_buf
[op_index
] = op_to_movi(op
);
1012 tmp
= do_constant_folding(op
, temps
[args
[1]].val
, 0);
1013 tcg_opt_gen_movi(gen_args
, args
[0], tmp
);
1029 CASE_OP_32_64(rotl
):
1030 CASE_OP_32_64(rotr
):
1031 CASE_OP_32_64(andc
):
1034 CASE_OP_32_64(nand
):
1036 CASE_OP_32_64(muluh
):
1037 CASE_OP_32_64(mulsh
):
1039 CASE_OP_32_64(divu
):
1041 CASE_OP_32_64(remu
):
1042 if (temps
[args
[1]].state
== TCG_TEMP_CONST
1043 && temps
[args
[2]].state
== TCG_TEMP_CONST
) {
1044 s
->gen_opc_buf
[op_index
] = op_to_movi(op
);
1045 tmp
= do_constant_folding(op
, temps
[args
[1]].val
,
1046 temps
[args
[2]].val
);
1047 tcg_opt_gen_movi(gen_args
, args
[0], tmp
);
1054 CASE_OP_32_64(deposit
):
1055 if (temps
[args
[1]].state
== TCG_TEMP_CONST
1056 && temps
[args
[2]].state
== TCG_TEMP_CONST
) {
1057 s
->gen_opc_buf
[op_index
] = op_to_movi(op
);
1058 tmp
= ((1ull << args
[4]) - 1);
1059 tmp
= (temps
[args
[1]].val
& ~(tmp
<< args
[3]))
1060 | ((temps
[args
[2]].val
& tmp
) << args
[3]);
1061 tcg_opt_gen_movi(gen_args
, args
[0], tmp
);
1068 CASE_OP_32_64(setcond
):
1069 tmp
= do_constant_folding_cond(op
, args
[1], args
[2], args
[3]);
1071 s
->gen_opc_buf
[op_index
] = op_to_movi(op
);
1072 tcg_opt_gen_movi(gen_args
, args
[0], tmp
);
1079 CASE_OP_32_64(brcond
):
1080 tmp
= do_constant_folding_cond(op
, args
[0], args
[1], args
[2]);
1083 reset_all_temps(nb_temps
);
1084 s
->gen_opc_buf
[op_index
] = INDEX_op_br
;
1085 gen_args
[0] = args
[3];
1088 s
->gen_opc_buf
[op_index
] = INDEX_op_nop
;
1095 CASE_OP_32_64(movcond
):
1096 tmp
= do_constant_folding_cond(op
, args
[1], args
[2], args
[5]);
1098 if (temps_are_copies(args
[0], args
[4-tmp
])) {
1099 s
->gen_opc_buf
[op_index
] = INDEX_op_nop
;
1100 } else if (temps
[args
[4-tmp
]].state
== TCG_TEMP_CONST
) {
1101 s
->gen_opc_buf
[op_index
] = op_to_movi(op
);
1102 tcg_opt_gen_movi(gen_args
, args
[0], temps
[args
[4-tmp
]].val
);
1105 s
->gen_opc_buf
[op_index
] = op_to_mov(op
);
1106 tcg_opt_gen_mov(s
, gen_args
, args
[0], args
[4-tmp
]);
1114 case INDEX_op_add2_i32
:
1115 case INDEX_op_sub2_i32
:
1116 if (temps
[args
[2]].state
== TCG_TEMP_CONST
1117 && temps
[args
[3]].state
== TCG_TEMP_CONST
1118 && temps
[args
[4]].state
== TCG_TEMP_CONST
1119 && temps
[args
[5]].state
== TCG_TEMP_CONST
) {
1120 uint32_t al
= temps
[args
[2]].val
;
1121 uint32_t ah
= temps
[args
[3]].val
;
1122 uint32_t bl
= temps
[args
[4]].val
;
1123 uint32_t bh
= temps
[args
[5]].val
;
1124 uint64_t a
= ((uint64_t)ah
<< 32) | al
;
1125 uint64_t b
= ((uint64_t)bh
<< 32) | bl
;
1128 if (op
== INDEX_op_add2_i32
) {
1134 /* We emit the extra nop when we emit the add2/sub2. */
1135 assert(s
->gen_opc_buf
[op_index
+ 1] == INDEX_op_nop
);
1139 s
->gen_opc_buf
[op_index
] = INDEX_op_movi_i32
;
1140 s
->gen_opc_buf
[++op_index
] = INDEX_op_movi_i32
;
1141 tcg_opt_gen_movi(&gen_args
[0], rl
, (uint32_t)a
);
1142 tcg_opt_gen_movi(&gen_args
[2], rh
, (uint32_t)(a
>> 32));
1149 case INDEX_op_mulu2_i32
:
1150 if (temps
[args
[2]].state
== TCG_TEMP_CONST
1151 && temps
[args
[3]].state
== TCG_TEMP_CONST
) {
1152 uint32_t a
= temps
[args
[2]].val
;
1153 uint32_t b
= temps
[args
[3]].val
;
1154 uint64_t r
= (uint64_t)a
* b
;
1157 /* We emit the extra nop when we emit the mulu2. */
1158 assert(s
->gen_opc_buf
[op_index
+ 1] == INDEX_op_nop
);
1162 s
->gen_opc_buf
[op_index
] = INDEX_op_movi_i32
;
1163 s
->gen_opc_buf
[++op_index
] = INDEX_op_movi_i32
;
1164 tcg_opt_gen_movi(&gen_args
[0], rl
, (uint32_t)r
);
1165 tcg_opt_gen_movi(&gen_args
[2], rh
, (uint32_t)(r
>> 32));
1172 case INDEX_op_brcond2_i32
:
1173 tmp
= do_constant_folding_cond2(&args
[0], &args
[2], args
[4]);
1176 reset_all_temps(nb_temps
);
1177 s
->gen_opc_buf
[op_index
] = INDEX_op_br
;
1178 gen_args
[0] = args
[5];
1181 s
->gen_opc_buf
[op_index
] = INDEX_op_nop
;
1183 } else if ((args
[4] == TCG_COND_LT
|| args
[4] == TCG_COND_GE
)
1184 && temps
[args
[2]].state
== TCG_TEMP_CONST
1185 && temps
[args
[3]].state
== TCG_TEMP_CONST
1186 && temps
[args
[2]].val
== 0
1187 && temps
[args
[3]].val
== 0) {
1188 /* Simplify LT/GE comparisons vs zero to a single compare
1189 vs the high word of the input. */
1190 reset_all_temps(nb_temps
);
1191 s
->gen_opc_buf
[op_index
] = INDEX_op_brcond_i32
;
1192 gen_args
[0] = args
[1];
1193 gen_args
[1] = args
[3];
1194 gen_args
[2] = args
[4];
1195 gen_args
[3] = args
[5];
1203 case INDEX_op_setcond2_i32
:
1204 tmp
= do_constant_folding_cond2(&args
[1], &args
[3], args
[5]);
1206 s
->gen_opc_buf
[op_index
] = INDEX_op_movi_i32
;
1207 tcg_opt_gen_movi(gen_args
, args
[0], tmp
);
1209 } else if ((args
[5] == TCG_COND_LT
|| args
[5] == TCG_COND_GE
)
1210 && temps
[args
[3]].state
== TCG_TEMP_CONST
1211 && temps
[args
[4]].state
== TCG_TEMP_CONST
1212 && temps
[args
[3]].val
== 0
1213 && temps
[args
[4]].val
== 0) {
1214 /* Simplify LT/GE comparisons vs zero to a single compare
1215 vs the high word of the input. */
1216 s
->gen_opc_buf
[op_index
] = INDEX_op_setcond_i32
;
1217 reset_temp(args
[0]);
1218 gen_args
[0] = args
[0];
1219 gen_args
[1] = args
[2];
1220 gen_args
[2] = args
[4];
1221 gen_args
[3] = args
[5];
1230 nb_call_args
= (args
[0] >> 16) + (args
[0] & 0xffff);
1231 if (!(args
[nb_call_args
+ 1] & (TCG_CALL_NO_READ_GLOBALS
|
1232 TCG_CALL_NO_WRITE_GLOBALS
))) {
1233 for (i
= 0; i
< nb_globals
; i
++) {
1237 for (i
= 0; i
< (args
[0] >> 16); i
++) {
1238 reset_temp(args
[i
+ 1]);
1240 i
= nb_call_args
+ 3;
1251 /* Default case: we know nothing about operation (or were unable
1252 to compute the operation result) so no propagation is done.
1253 We trash everything if the operation is the end of a basic
1254 block, otherwise we only trash the output args. "mask" is
1255 the non-zero bits mask for the first output arg. */
1256 if (def
->flags
& TCG_OPF_BB_END
) {
1257 reset_all_temps(nb_temps
);
1259 for (i
= 0; i
< def
->nb_oargs
; i
++) {
1260 reset_temp(args
[i
]);
1261 /* Save the corresponding known-zero bits mask for the
1262 first output argument (only one supported so far). */
1264 temps
[args
[i
]].mask
= mask
;
1268 for (i
= 0; i
< def
->nb_args
; i
++) {
1269 gen_args
[i
] = args
[i
];
1271 args
+= def
->nb_args
;
1272 gen_args
+= def
->nb_args
;
1280 TCGArg
*tcg_optimize(TCGContext
*s
, uint16_t *tcg_opc_ptr
,
1281 TCGArg
*args
, TCGOpDef
*tcg_op_defs
)
1284 res
= tcg_constant_folding(s
, tcg_opc_ptr
, args
, tcg_op_defs
);