]>
git.proxmox.com Git - qemu.git/blob - tcg/optimize.c
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)
46 struct tcg_temp_info
{
53 static struct tcg_temp_info temps
[TCG_MAX_TEMPS
];
55 /* Reset TEMP's state to TCG_TEMP_ANY. If TEMP was a representative of some
56 class of equivalent temp's, a new representative should be chosen in this
58 static void reset_temp(TCGArg temp
, int nb_temps
, int nb_globals
)
61 TCGArg new_base
= (TCGArg
)-1;
62 if (temps
[temp
].state
== TCG_TEMP_HAS_COPY
) {
63 for (i
= temps
[temp
].next_copy
; i
!= temp
; i
= temps
[i
].next_copy
) {
64 if (i
>= nb_globals
) {
65 temps
[i
].state
= TCG_TEMP_HAS_COPY
;
70 for (i
= temps
[temp
].next_copy
; i
!= temp
; i
= temps
[i
].next_copy
) {
71 if (new_base
== (TCGArg
)-1) {
72 temps
[i
].state
= TCG_TEMP_ANY
;
74 temps
[i
].val
= new_base
;
77 temps
[temps
[temp
].next_copy
].prev_copy
= temps
[temp
].prev_copy
;
78 temps
[temps
[temp
].prev_copy
].next_copy
= temps
[temp
].next_copy
;
79 } else if (temps
[temp
].state
== TCG_TEMP_COPY
) {
80 temps
[temps
[temp
].next_copy
].prev_copy
= temps
[temp
].prev_copy
;
81 temps
[temps
[temp
].prev_copy
].next_copy
= temps
[temp
].next_copy
;
82 new_base
= temps
[temp
].val
;
84 temps
[temp
].state
= TCG_TEMP_ANY
;
85 if (new_base
!= (TCGArg
)-1 && temps
[new_base
].next_copy
== new_base
) {
86 temps
[new_base
].state
= TCG_TEMP_ANY
;
90 static int op_bits(TCGOpcode op
)
92 const TCGOpDef
*def
= &tcg_op_defs
[op
];
93 return def
->flags
& TCG_OPF_64BIT
? 64 : 32;
96 static TCGOpcode
op_to_movi(TCGOpcode op
)
98 switch (op_bits(op
)) {
100 return INDEX_op_movi_i32
;
102 return INDEX_op_movi_i64
;
104 fprintf(stderr
, "op_to_movi: unexpected return value of "
105 "function op_bits.\n");
110 static void tcg_opt_gen_mov(TCGContext
*s
, TCGArg
*gen_args
, TCGArg dst
,
111 TCGArg src
, int nb_temps
, int nb_globals
)
113 reset_temp(dst
, nb_temps
, nb_globals
);
114 assert(temps
[src
].state
!= TCG_TEMP_COPY
);
115 /* Don't try to copy if one of temps is a global or either one
116 is local and another is register */
117 if (src
>= nb_globals
&& dst
>= nb_globals
&&
118 tcg_arg_is_local(s
, src
) == tcg_arg_is_local(s
, dst
)) {
119 assert(temps
[src
].state
!= TCG_TEMP_CONST
);
120 if (temps
[src
].state
!= TCG_TEMP_HAS_COPY
) {
121 temps
[src
].state
= TCG_TEMP_HAS_COPY
;
122 temps
[src
].next_copy
= src
;
123 temps
[src
].prev_copy
= src
;
125 temps
[dst
].state
= TCG_TEMP_COPY
;
126 temps
[dst
].val
= src
;
127 temps
[dst
].next_copy
= temps
[src
].next_copy
;
128 temps
[dst
].prev_copy
= src
;
129 temps
[temps
[dst
].next_copy
].prev_copy
= dst
;
130 temps
[src
].next_copy
= dst
;
136 static void tcg_opt_gen_movi(TCGArg
*gen_args
, TCGArg dst
, TCGArg val
,
137 int nb_temps
, int nb_globals
)
139 reset_temp(dst
, nb_temps
, nb_globals
);
140 temps
[dst
].state
= TCG_TEMP_CONST
;
141 temps
[dst
].val
= val
;
146 static TCGOpcode
op_to_mov(TCGOpcode op
)
148 switch (op_bits(op
)) {
150 return INDEX_op_mov_i32
;
152 return INDEX_op_mov_i64
;
154 fprintf(stderr
, "op_to_mov: unexpected return value of "
155 "function op_bits.\n");
160 static TCGArg
do_constant_folding_2(TCGOpcode op
, TCGArg x
, TCGArg y
)
181 case INDEX_op_shl_i32
:
182 return (uint32_t)x
<< (uint32_t)y
;
184 case INDEX_op_shl_i64
:
185 return (uint64_t)x
<< (uint64_t)y
;
187 case INDEX_op_shr_i32
:
188 return (uint32_t)x
>> (uint32_t)y
;
190 case INDEX_op_shr_i64
:
191 return (uint64_t)x
>> (uint64_t)y
;
193 case INDEX_op_sar_i32
:
194 return (int32_t)x
>> (int32_t)y
;
196 case INDEX_op_sar_i64
:
197 return (int64_t)x
>> (int64_t)y
;
199 case INDEX_op_rotr_i32
:
200 x
= ((uint32_t)x
<< (32 - y
)) | ((uint32_t)x
>> y
);
203 case INDEX_op_rotr_i64
:
204 x
= ((uint64_t)x
<< (64 - y
)) | ((uint64_t)x
>> y
);
207 case INDEX_op_rotl_i32
:
208 x
= ((uint32_t)x
<< y
) | ((uint32_t)x
>> (32 - y
));
211 case INDEX_op_rotl_i64
:
212 x
= ((uint64_t)x
<< y
) | ((uint64_t)x
>> (64 - y
));
236 CASE_OP_32_64(ext8s
):
239 CASE_OP_32_64(ext16s
):
242 CASE_OP_32_64(ext8u
):
245 CASE_OP_32_64(ext16u
):
248 case INDEX_op_ext32s_i64
:
251 case INDEX_op_ext32u_i64
:
256 "Unrecognized operation %d in do_constant_folding.\n", op
);
261 static TCGArg
do_constant_folding(TCGOpcode op
, TCGArg x
, TCGArg y
)
263 TCGArg res
= do_constant_folding_2(op
, x
, y
);
264 if (op_bits(op
) == 32) {
270 /* Propagate constants and copies, fold constant expressions. */
271 static TCGArg
*tcg_constant_folding(TCGContext
*s
, uint16_t *tcg_opc_ptr
,
272 TCGArg
*args
, TCGOpDef
*tcg_op_defs
)
274 int i
, nb_ops
, op_index
, nb_temps
, nb_globals
, nb_call_args
;
279 /* Array VALS has an element for each temp.
280 If this temp holds a constant then its value is kept in VALS' element.
281 If this temp is a copy of other ones then this equivalence class'
282 representative is kept in VALS' element.
283 If this temp is neither copy nor constant then corresponding VALS'
284 element is unused. */
286 nb_temps
= s
->nb_temps
;
287 nb_globals
= s
->nb_globals
;
288 memset(temps
, 0, nb_temps
* sizeof(struct tcg_temp_info
));
290 nb_ops
= tcg_opc_ptr
- gen_opc_buf
;
292 for (op_index
= 0; op_index
< nb_ops
; op_index
++) {
293 op
= gen_opc_buf
[op_index
];
294 def
= &tcg_op_defs
[op
];
295 /* Do copy propagation */
296 if (!(def
->flags
& (TCG_OPF_CALL_CLOBBER
| TCG_OPF_SIDE_EFFECTS
))) {
297 assert(op
!= INDEX_op_call
);
298 for (i
= def
->nb_oargs
; i
< def
->nb_oargs
+ def
->nb_iargs
; i
++) {
299 if (temps
[args
[i
]].state
== TCG_TEMP_COPY
) {
300 args
[i
] = temps
[args
[i
]].val
;
305 /* For commutative operations make constant second argument */
315 if (temps
[args
[1]].state
== TCG_TEMP_CONST
) {
321 CASE_OP_32_64(brcond
):
322 if (temps
[args
[0]].state
== TCG_TEMP_CONST
323 && temps
[args
[1]].state
!= TCG_TEMP_CONST
) {
327 args
[2] = tcg_swap_cond(args
[2]);
330 CASE_OP_32_64(setcond
):
331 if (temps
[args
[1]].state
== TCG_TEMP_CONST
332 && temps
[args
[2]].state
!= TCG_TEMP_CONST
) {
336 args
[3] = tcg_swap_cond(args
[3]);
343 /* Simplify expressions for "shift/rot r, 0, a => movi r, 0" */
350 if (temps
[args
[1]].state
== TCG_TEMP_CONST
351 && temps
[args
[1]].val
== 0) {
352 gen_opc_buf
[op_index
] = op_to_movi(op
);
353 tcg_opt_gen_movi(gen_args
, args
[0], 0, nb_temps
, nb_globals
);
363 /* Simplify expression for "op r, a, 0 => mov r, a" cases */
374 if (temps
[args
[1]].state
== TCG_TEMP_CONST
) {
375 /* Proceed with possible constant folding. */
378 if (temps
[args
[2]].state
== TCG_TEMP_CONST
379 && temps
[args
[2]].val
== 0) {
380 if ((temps
[args
[0]].state
== TCG_TEMP_COPY
381 && temps
[args
[0]].val
== args
[1])
382 || args
[0] == args
[1]) {
384 gen_opc_buf
[op_index
] = INDEX_op_nop
;
386 gen_opc_buf
[op_index
] = op_to_mov(op
);
387 tcg_opt_gen_mov(s
, gen_args
, args
[0], args
[1],
388 nb_temps
, nb_globals
);
399 /* Simplify expression for "op r, a, 0 => movi r, 0" cases */
403 if ((temps
[args
[2]].state
== TCG_TEMP_CONST
404 && temps
[args
[2]].val
== 0)) {
405 gen_opc_buf
[op_index
] = op_to_movi(op
);
406 tcg_opt_gen_movi(gen_args
, args
[0], 0, nb_temps
, nb_globals
);
416 /* Simplify expression for "op r, a, a => mov r, a" cases */
420 if (args
[1] == args
[2]) {
421 if (args
[1] == args
[0]) {
423 gen_opc_buf
[op_index
] = INDEX_op_nop
;
425 gen_opc_buf
[op_index
] = op_to_mov(op
);
426 tcg_opt_gen_mov(s
, gen_args
, args
[0], args
[1], nb_temps
,
438 /* Propagate constants through copy operations and do constant
439 folding. Constants will be substituted to arguments by register
440 allocator where needed and possible. Also detect copies. */
443 if ((temps
[args
[1]].state
== TCG_TEMP_COPY
444 && temps
[args
[1]].val
== args
[0])
445 || args
[0] == args
[1]) {
447 gen_opc_buf
[op_index
] = INDEX_op_nop
;
450 if (temps
[args
[1]].state
!= TCG_TEMP_CONST
) {
451 tcg_opt_gen_mov(s
, gen_args
, args
[0], args
[1],
452 nb_temps
, nb_globals
);
457 /* Source argument is constant. Rewrite the operation and
458 let movi case handle it. */
460 gen_opc_buf
[op_index
] = op
;
461 args
[1] = temps
[args
[1]].val
;
464 tcg_opt_gen_movi(gen_args
, args
[0], args
[1], nb_temps
, nb_globals
);
470 CASE_OP_32_64(ext8s
):
471 CASE_OP_32_64(ext8u
):
472 CASE_OP_32_64(ext16s
):
473 CASE_OP_32_64(ext16u
):
474 case INDEX_op_ext32s_i64
:
475 case INDEX_op_ext32u_i64
:
476 if (temps
[args
[1]].state
== TCG_TEMP_CONST
) {
477 gen_opc_buf
[op_index
] = op_to_movi(op
);
478 tmp
= do_constant_folding(op
, temps
[args
[1]].val
, 0);
479 tcg_opt_gen_movi(gen_args
, args
[0], tmp
, nb_temps
, nb_globals
);
484 reset_temp(args
[0], nb_temps
, nb_globals
);
485 gen_args
[0] = args
[0];
486 gen_args
[1] = args
[1];
507 if (temps
[args
[1]].state
== TCG_TEMP_CONST
508 && temps
[args
[2]].state
== TCG_TEMP_CONST
) {
509 gen_opc_buf
[op_index
] = op_to_movi(op
);
510 tmp
= do_constant_folding(op
, temps
[args
[1]].val
,
512 tcg_opt_gen_movi(gen_args
, args
[0], tmp
, nb_temps
, nb_globals
);
517 reset_temp(args
[0], nb_temps
, nb_globals
);
518 gen_args
[0] = args
[0];
519 gen_args
[1] = args
[1];
520 gen_args
[2] = args
[2];
526 nb_call_args
= (args
[0] >> 16) + (args
[0] & 0xffff);
527 if (!(args
[nb_call_args
+ 1] & (TCG_CALL_CONST
| TCG_CALL_PURE
))) {
528 for (i
= 0; i
< nb_globals
; i
++) {
529 reset_temp(i
, nb_temps
, nb_globals
);
532 for (i
= 0; i
< (args
[0] >> 16); i
++) {
533 reset_temp(args
[i
+ 1], nb_temps
, nb_globals
);
535 i
= nb_call_args
+ 3;
543 case INDEX_op_set_label
:
546 CASE_OP_32_64(brcond
):
547 memset(temps
, 0, nb_temps
* sizeof(struct tcg_temp_info
));
548 for (i
= 0; i
< def
->nb_args
; i
++) {
555 /* Default case: we do know nothing about operation so no
556 propagation is done. We only trash output args. */
557 for (i
= 0; i
< def
->nb_oargs
; i
++) {
558 reset_temp(args
[i
], nb_temps
, nb_globals
);
560 for (i
= 0; i
< def
->nb_args
; i
++) {
561 gen_args
[i
] = args
[i
];
563 args
+= def
->nb_args
;
564 gen_args
+= def
->nb_args
;
572 TCGArg
*tcg_optimize(TCGContext
*s
, uint16_t *tcg_opc_ptr
,
573 TCGArg
*args
, TCGOpDef
*tcg_op_defs
)
576 res
= tcg_constant_folding(s
, tcg_opc_ptr
, args
, tcg_op_defs
);