2 * Optimizations for Tiny Code Generator for QEMU
4 * Copyright (c) 2010 Samsung Electronics.
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 {
51 static struct tcg_temp_info temps[TCG_MAX_TEMPS];
53 /* Reset TEMP's state to TCG_TEMP_UNDEF. If TEMP only had one copy, remove
54 the copy flag from the left temp. */
55 static void reset_temp(TCGArg temp)
57 if (temps[temp].state == TCG_TEMP_COPY) {
58 if (temps[temp].prev_copy == temps[temp].next_copy) {
59 temps[temps[temp].next_copy].state = TCG_TEMP_UNDEF;
61 temps[temps[temp].next_copy].prev_copy = temps[temp].prev_copy;
62 temps[temps[temp].prev_copy].next_copy = temps[temp].next_copy;
65 temps[temp].state = TCG_TEMP_UNDEF;
68 static int op_bits(TCGOpcode op)
70 const TCGOpDef *def = &tcg_op_defs[op];
71 return def->flags & TCG_OPF_64BIT ? 64 : 32;
74 static TCGOpcode op_to_movi(TCGOpcode op)
76 switch (op_bits(op)) {
78 return INDEX_op_movi_i32;
80 return INDEX_op_movi_i64;
82 fprintf(stderr, "op_to_movi: unexpected return value of "
83 "function op_bits.\n");
88 static TCGArg find_better_copy(TCGContext *s, TCGArg temp)
92 /* If this is already a global, we can't do better. */
93 if (temp < s->nb_globals) {
97 /* Search for a global first. */
98 for (i = temps[temp].next_copy ; i != temp ; i = temps[i].next_copy) {
99 if (i < s->nb_globals) {
104 /* If it is a temp, search for a temp local. */
105 if (!s->temps[temp].temp_local) {
106 for (i = temps[temp].next_copy ; i != temp ; i = temps[i].next_copy) {
107 if (s->temps[i].temp_local) {
113 /* Failure to find a better representation, return the same temp. */
117 static bool temps_are_copies(TCGArg arg1, TCGArg arg2)
125 if (temps[arg1].state != TCG_TEMP_COPY
126 || temps[arg2].state != TCG_TEMP_COPY) {
130 for (i = temps[arg1].next_copy ; i != arg1 ; i = temps[i].next_copy) {
139 static void tcg_opt_gen_mov(TCGContext *s, TCGArg *gen_args,
140 TCGArg dst, TCGArg src)
143 assert(temps[src].state != TCG_TEMP_CONST);
145 if (s->temps[src].type == s->temps[dst].type) {
146 if (temps[src].state != TCG_TEMP_COPY) {
147 temps[src].state = TCG_TEMP_COPY;
148 temps[src].next_copy = src;
149 temps[src].prev_copy = src;
151 temps[dst].state = TCG_TEMP_COPY;
152 temps[dst].next_copy = temps[src].next_copy;
153 temps[dst].prev_copy = src;
154 temps[temps[dst].next_copy].prev_copy = dst;
155 temps[src].next_copy = dst;
162 static void tcg_opt_gen_movi(TCGArg *gen_args, TCGArg dst, TCGArg val)
165 temps[dst].state = TCG_TEMP_CONST;
166 temps[dst].val = val;
171 static TCGOpcode op_to_mov(TCGOpcode op)
173 switch (op_bits(op)) {
175 return INDEX_op_mov_i32;
177 return INDEX_op_mov_i64;
179 fprintf(stderr, "op_to_mov: unexpected return value of "
180 "function op_bits.\n");
185 static TCGArg do_constant_folding_2(TCGOpcode op, TCGArg x, TCGArg y)
206 case INDEX_op_shl_i32:
207 return (uint32_t)x << (uint32_t)y;
209 case INDEX_op_shl_i64:
210 return (uint64_t)x << (uint64_t)y;
212 case INDEX_op_shr_i32:
213 return (uint32_t)x >> (uint32_t)y;
215 case INDEX_op_shr_i64:
216 return (uint64_t)x >> (uint64_t)y;
218 case INDEX_op_sar_i32:
219 return (int32_t)x >> (int32_t)y;
221 case INDEX_op_sar_i64:
222 return (int64_t)x >> (int64_t)y;
224 case INDEX_op_rotr_i32:
225 x = ((uint32_t)x << (32 - y)) | ((uint32_t)x >> y);
228 case INDEX_op_rotr_i64:
229 x = ((uint64_t)x << (64 - y)) | ((uint64_t)x >> y);
232 case INDEX_op_rotl_i32:
233 x = ((uint32_t)x << y) | ((uint32_t)x >> (32 - y));
236 case INDEX_op_rotl_i64:
237 x = ((uint64_t)x << y) | ((uint64_t)x >> (64 - y));
261 CASE_OP_32_64(ext8s):
264 CASE_OP_32_64(ext16s):
267 CASE_OP_32_64(ext8u):
270 CASE_OP_32_64(ext16u):
273 case INDEX_op_ext32s_i64:
276 case INDEX_op_ext32u_i64:
281 "Unrecognized operation %d in do_constant_folding.\n", op);
286 static TCGArg do_constant_folding(TCGOpcode op, TCGArg x, TCGArg y)
288 TCGArg res = do_constant_folding_2(op, x, y);
289 if (op_bits(op) == 32) {
295 /* Return 2 if the condition can't be simplified, and the result
296 of the condition (0 or 1) if it can */
297 static TCGArg do_constant_folding_cond(TCGOpcode op, TCGArg x,
300 if (temps[x].state == TCG_TEMP_CONST && temps[y].state == TCG_TEMP_CONST) {
301 switch (op_bits(op)) {
305 return (uint32_t)temps[x].val == (uint32_t)temps[y].val;
307 return (uint32_t)temps[x].val != (uint32_t)temps[y].val;
309 return (int32_t)temps[x].val < (int32_t)temps[y].val;
311 return (int32_t)temps[x].val >= (int32_t)temps[y].val;
313 return (int32_t)temps[x].val <= (int32_t)temps[y].val;
315 return (int32_t)temps[x].val > (int32_t)temps[y].val;
317 return (uint32_t)temps[x].val < (uint32_t)temps[y].val;
319 return (uint32_t)temps[x].val >= (uint32_t)temps[y].val;
321 return (uint32_t)temps[x].val <= (uint32_t)temps[y].val;
323 return (uint32_t)temps[x].val > (uint32_t)temps[y].val;
329 return (uint64_t)temps[x].val == (uint64_t)temps[y].val;
331 return (uint64_t)temps[x].val != (uint64_t)temps[y].val;
333 return (int64_t)temps[x].val < (int64_t)temps[y].val;
335 return (int64_t)temps[x].val >= (int64_t)temps[y].val;
337 return (int64_t)temps[x].val <= (int64_t)temps[y].val;
339 return (int64_t)temps[x].val > (int64_t)temps[y].val;
341 return (uint64_t)temps[x].val < (uint64_t)temps[y].val;
343 return (uint64_t)temps[x].val >= (uint64_t)temps[y].val;
345 return (uint64_t)temps[x].val <= (uint64_t)temps[y].val;
347 return (uint64_t)temps[x].val > (uint64_t)temps[y].val;
351 } else if (temps_are_copies(x, y)) {
366 } else if (temps[y].state == TCG_TEMP_CONST && temps[y].val == 0) {
380 "Unrecognized bitness %d or condition %d in "
381 "do_constant_folding_cond.\n", op_bits(op), c);
385 /* Propagate constants and copies, fold constant expressions. */
386 static TCGArg *tcg_constant_folding(TCGContext *s, uint16_t *tcg_opc_ptr,
387 TCGArg *args, TCGOpDef *tcg_op_defs)
389 int i, nb_ops, op_index, nb_temps, nb_globals, nb_call_args;
396 /* Array VALS has an element for each temp.
397 If this temp holds a constant then its value is kept in VALS' element.
398 If this temp is a copy of other ones then the other copies are
399 available through the doubly linked circular list. */
401 nb_temps = s->nb_temps;
402 nb_globals = s->nb_globals;
403 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
405 nb_ops = tcg_opc_ptr - gen_opc_buf;
407 for (op_index = 0; op_index < nb_ops; op_index++) {
408 op = gen_opc_buf[op_index];
409 def = &tcg_op_defs[op];
410 /* Do copy propagation */
411 if (op == INDEX_op_call) {
412 int nb_oargs = args[0] >> 16;
413 int nb_iargs = args[0] & 0xffff;
414 for (i = nb_oargs + 1; i < nb_oargs + nb_iargs + 1; i++) {
415 if (temps[args[i]].state == TCG_TEMP_COPY) {
416 args[i] = find_better_copy(s, args[i]);
420 for (i = def->nb_oargs; i < def->nb_oargs + def->nb_iargs; i++) {
421 if (temps[args[i]].state == TCG_TEMP_COPY) {
422 args[i] = find_better_copy(s, args[i]);
427 /* For commutative operations make constant second argument */
437 if (temps[args[1]].state == TCG_TEMP_CONST) {
443 CASE_OP_32_64(brcond):
444 if (temps[args[0]].state == TCG_TEMP_CONST
445 && temps[args[1]].state != TCG_TEMP_CONST) {
449 args[2] = tcg_swap_cond(args[2]);
452 CASE_OP_32_64(setcond):
453 if (temps[args[1]].state == TCG_TEMP_CONST
454 && temps[args[2]].state != TCG_TEMP_CONST) {
458 args[3] = tcg_swap_cond(args[3]);
461 CASE_OP_32_64(movcond):
463 if (temps[args[1]].state == TCG_TEMP_CONST
464 && temps[args[2]].state != TCG_TEMP_CONST) {
468 cond = tcg_swap_cond(cond);
470 /* For movcond, we canonicalize the "false" input reg to match
471 the destination reg so that the tcg backend can implement
472 a "move if true" operation. */
473 if (args[0] == args[3]) {
477 cond = tcg_invert_cond(cond);
484 /* Simplify expressions for "shift/rot r, 0, a => movi r, 0" */
491 if (temps[args[1]].state == TCG_TEMP_CONST
492 && temps[args[1]].val == 0) {
493 gen_opc_buf[op_index] = op_to_movi(op);
494 tcg_opt_gen_movi(gen_args, args[0], 0);
504 /* Simplify expression for "op r, a, 0 => mov r, a" cases */
515 if (temps[args[1]].state == TCG_TEMP_CONST) {
516 /* Proceed with possible constant folding. */
519 if (temps[args[2]].state == TCG_TEMP_CONST
520 && temps[args[2]].val == 0) {
521 if (temps_are_copies(args[0], args[1])) {
522 gen_opc_buf[op_index] = INDEX_op_nop;
524 gen_opc_buf[op_index] = op_to_mov(op);
525 tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
536 /* Simplify expression for "op r, a, 0 => movi r, 0" cases */
540 if ((temps[args[2]].state == TCG_TEMP_CONST
541 && temps[args[2]].val == 0)) {
542 gen_opc_buf[op_index] = op_to_movi(op);
543 tcg_opt_gen_movi(gen_args, args[0], 0);
553 /* Simplify expression for "op r, a, a => mov r, a" cases */
557 if (temps_are_copies(args[1], args[2])) {
558 if (temps_are_copies(args[0], args[1])) {
559 gen_opc_buf[op_index] = INDEX_op_nop;
561 gen_opc_buf[op_index] = op_to_mov(op);
562 tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
573 /* Simplify expression for "op r, a, a => movi r, 0" cases */
577 if (temps_are_copies(args[1], args[2])) {
578 gen_opc_buf[op_index] = op_to_movi(op);
579 tcg_opt_gen_movi(gen_args, args[0], 0);
589 /* Propagate constants through copy operations and do constant
590 folding. Constants will be substituted to arguments by register
591 allocator where needed and possible. Also detect copies. */
594 if (temps_are_copies(args[0], args[1])) {
596 gen_opc_buf[op_index] = INDEX_op_nop;
599 if (temps[args[1]].state != TCG_TEMP_CONST) {
600 tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
605 /* Source argument is constant. Rewrite the operation and
606 let movi case handle it. */
608 gen_opc_buf[op_index] = op;
609 args[1] = temps[args[1]].val;
612 tcg_opt_gen_movi(gen_args, args[0], args[1]);
618 CASE_OP_32_64(ext8s):
619 CASE_OP_32_64(ext8u):
620 CASE_OP_32_64(ext16s):
621 CASE_OP_32_64(ext16u):
622 case INDEX_op_ext32s_i64:
623 case INDEX_op_ext32u_i64:
624 if (temps[args[1]].state == TCG_TEMP_CONST) {
625 gen_opc_buf[op_index] = op_to_movi(op);
626 tmp = do_constant_folding(op, temps[args[1]].val, 0);
627 tcg_opt_gen_movi(gen_args, args[0], tmp);
630 gen_args[0] = args[0];
631 gen_args[1] = args[1];
652 if (temps[args[1]].state == TCG_TEMP_CONST
653 && temps[args[2]].state == TCG_TEMP_CONST) {
654 gen_opc_buf[op_index] = op_to_movi(op);
655 tmp = do_constant_folding(op, temps[args[1]].val,
657 tcg_opt_gen_movi(gen_args, args[0], tmp);
661 gen_args[0] = args[0];
662 gen_args[1] = args[1];
663 gen_args[2] = args[2];
668 CASE_OP_32_64(setcond):
669 tmp = do_constant_folding_cond(op, args[1], args[2], args[3]);
671 gen_opc_buf[op_index] = op_to_movi(op);
672 tcg_opt_gen_movi(gen_args, args[0], tmp);
676 gen_args[0] = args[0];
677 gen_args[1] = args[1];
678 gen_args[2] = args[2];
679 gen_args[3] = args[3];
684 CASE_OP_32_64(brcond):
685 tmp = do_constant_folding_cond(op, args[0], args[1], args[2]);
688 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
689 gen_opc_buf[op_index] = INDEX_op_br;
690 gen_args[0] = args[3];
693 gen_opc_buf[op_index] = INDEX_op_nop;
696 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
698 gen_args[0] = args[0];
699 gen_args[1] = args[1];
700 gen_args[2] = args[2];
701 gen_args[3] = args[3];
706 CASE_OP_32_64(movcond):
707 tmp = do_constant_folding_cond(op, args[1], args[2], args[5]);
709 if (temps_are_copies(args[0], args[4-tmp])) {
710 gen_opc_buf[op_index] = INDEX_op_nop;
711 } else if (temps[args[4-tmp]].state == TCG_TEMP_CONST) {
712 gen_opc_buf[op_index] = op_to_movi(op);
713 tcg_opt_gen_movi(gen_args, args[0], temps[args[4-tmp]].val);
716 gen_opc_buf[op_index] = op_to_mov(op);
717 tcg_opt_gen_mov(s, gen_args, args[0], args[4-tmp]);
722 gen_args[0] = args[0];
723 gen_args[1] = args[1];
724 gen_args[2] = args[2];
725 gen_args[3] = args[3];
726 gen_args[4] = args[4];
727 gen_args[5] = args[5];
733 nb_call_args = (args[0] >> 16) + (args[0] & 0xffff);
734 if (!(args[nb_call_args + 1] & (TCG_CALL_CONST | TCG_CALL_PURE))) {
735 for (i = 0; i < nb_globals; i++) {
739 for (i = 0; i < (args[0] >> 16); i++) {
740 reset_temp(args[i + 1]);
742 i = nb_call_args + 3;
751 /* Default case: we do know nothing about operation so no
752 propagation is done. We trash everything if the operation
753 is the end of a basic block, otherwise we only trash the
755 if (def->flags & TCG_OPF_BB_END) {
756 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
758 for (i = 0; i < def->nb_oargs; i++) {
762 for (i = 0; i < def->nb_args; i++) {
763 gen_args[i] = args[i];
765 args += def->nb_args;
766 gen_args += def->nb_args;
774 TCGArg *tcg_optimize(TCGContext *s, uint16_t *tcg_opc_ptr,
775 TCGArg *args, TCGOpDef *tcg_op_defs)
778 res = tcg_constant_folding(s, tcg_opc_ptr, args, tcg_op_defs);