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 static bool do_constant_folding_cond_32(uint32_t x, uint32_t y, TCGCond c)
303 return (int32_t)x < (int32_t)y;
305 return (int32_t)x >= (int32_t)y;
307 return (int32_t)x <= (int32_t)y;
309 return (int32_t)x > (int32_t)y;
323 static bool do_constant_folding_cond_64(uint64_t x, uint64_t y, TCGCond c)
331 return (int64_t)x < (int64_t)y;
333 return (int64_t)x >= (int64_t)y;
335 return (int64_t)x <= (int64_t)y;
337 return (int64_t)x > (int64_t)y;
351 static bool do_constant_folding_cond_eq(TCGCond c)
371 /* Return 2 if the condition can't be simplified, and the result
372 of the condition (0 or 1) if it can */
373 static TCGArg do_constant_folding_cond(TCGOpcode op, TCGArg x,
376 if (temps[x].state == TCG_TEMP_CONST && temps[y].state == TCG_TEMP_CONST) {
377 switch (op_bits(op)) {
379 return do_constant_folding_cond_32(temps[x].val, temps[y].val, c);
381 return do_constant_folding_cond_64(temps[x].val, temps[y].val, c);
385 } else if (temps_are_copies(x, y)) {
386 return do_constant_folding_cond_eq(c);
387 } else if (temps[y].state == TCG_TEMP_CONST && temps[y].val == 0) {
401 static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2)
403 TCGArg a1 = *p1, a2 = *p2;
405 sum += temps[a1].state == TCG_TEMP_CONST;
406 sum -= temps[a2].state == TCG_TEMP_CONST;
408 /* Prefer the constant in second argument, and then the form
409 op a, a, b, which is better handled on non-RISC hosts. */
410 if (sum > 0 || (sum == 0 && dest == a2)) {
418 static bool swap_commutative2(TCGArg *p1, TCGArg *p2)
421 sum += temps[p1[0]].state == TCG_TEMP_CONST;
422 sum += temps[p1[1]].state == TCG_TEMP_CONST;
423 sum -= temps[p2[0]].state == TCG_TEMP_CONST;
424 sum -= temps[p2[1]].state == TCG_TEMP_CONST;
427 t = p1[0], p1[0] = p2[0], p2[0] = t;
428 t = p1[1], p1[1] = p2[1], p2[1] = t;
434 /* Propagate constants and copies, fold constant expressions. */
435 static TCGArg *tcg_constant_folding(TCGContext *s, uint16_t *tcg_opc_ptr,
436 TCGArg *args, TCGOpDef *tcg_op_defs)
438 int i, nb_ops, op_index, nb_temps, nb_globals, nb_call_args;
444 /* Array VALS has an element for each temp.
445 If this temp holds a constant then its value is kept in VALS' element.
446 If this temp is a copy of other ones then the other copies are
447 available through the doubly linked circular list. */
449 nb_temps = s->nb_temps;
450 nb_globals = s->nb_globals;
451 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
453 nb_ops = tcg_opc_ptr - gen_opc_buf;
455 for (op_index = 0; op_index < nb_ops; op_index++) {
456 op = gen_opc_buf[op_index];
457 def = &tcg_op_defs[op];
458 /* Do copy propagation */
459 if (op == INDEX_op_call) {
460 int nb_oargs = args[0] >> 16;
461 int nb_iargs = args[0] & 0xffff;
462 for (i = nb_oargs + 1; i < nb_oargs + nb_iargs + 1; i++) {
463 if (temps[args[i]].state == TCG_TEMP_COPY) {
464 args[i] = find_better_copy(s, args[i]);
468 for (i = def->nb_oargs; i < def->nb_oargs + def->nb_iargs; i++) {
469 if (temps[args[i]].state == TCG_TEMP_COPY) {
470 args[i] = find_better_copy(s, args[i]);
475 /* For commutative operations make constant second argument */
485 swap_commutative(args[0], &args[1], &args[2]);
487 CASE_OP_32_64(brcond):
488 if (swap_commutative(-1, &args[0], &args[1])) {
489 args[2] = tcg_swap_cond(args[2]);
492 CASE_OP_32_64(setcond):
493 if (swap_commutative(args[0], &args[1], &args[2])) {
494 args[3] = tcg_swap_cond(args[3]);
497 CASE_OP_32_64(movcond):
498 if (swap_commutative(-1, &args[1], &args[2])) {
499 args[5] = tcg_swap_cond(args[5]);
501 /* For movcond, we canonicalize the "false" input reg to match
502 the destination reg so that the tcg backend can implement
503 a "move if true" operation. */
504 if (swap_commutative(args[0], &args[4], &args[3])) {
505 args[5] = tcg_invert_cond(args[5]);
508 case INDEX_op_add2_i32:
509 swap_commutative(args[0], &args[2], &args[4]);
510 swap_commutative(args[1], &args[3], &args[5]);
512 case INDEX_op_brcond2_i32:
513 if (swap_commutative2(&args[0], &args[2])) {
514 args[4] = tcg_swap_cond(args[4]);
517 case INDEX_op_setcond2_i32:
518 if (swap_commutative2(&args[1], &args[3])) {
519 args[5] = tcg_swap_cond(args[5]);
526 /* Simplify expressions for "shift/rot r, 0, a => movi r, 0" */
533 if (temps[args[1]].state == TCG_TEMP_CONST
534 && temps[args[1]].val == 0) {
535 gen_opc_buf[op_index] = op_to_movi(op);
536 tcg_opt_gen_movi(gen_args, args[0], 0);
546 /* Simplify expression for "op r, a, 0 => mov r, a" cases */
557 if (temps[args[1]].state == TCG_TEMP_CONST) {
558 /* Proceed with possible constant folding. */
561 if (temps[args[2]].state == TCG_TEMP_CONST
562 && temps[args[2]].val == 0) {
563 if (temps_are_copies(args[0], args[1])) {
564 gen_opc_buf[op_index] = INDEX_op_nop;
566 gen_opc_buf[op_index] = op_to_mov(op);
567 tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
578 /* Simplify expression for "op r, a, 0 => movi r, 0" cases */
582 if ((temps[args[2]].state == TCG_TEMP_CONST
583 && temps[args[2]].val == 0)) {
584 gen_opc_buf[op_index] = op_to_movi(op);
585 tcg_opt_gen_movi(gen_args, args[0], 0);
595 /* Simplify expression for "op r, a, a => mov r, a" cases */
599 if (temps_are_copies(args[1], args[2])) {
600 if (temps_are_copies(args[0], args[1])) {
601 gen_opc_buf[op_index] = INDEX_op_nop;
603 gen_opc_buf[op_index] = op_to_mov(op);
604 tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
615 /* Simplify expression for "op r, a, a => movi r, 0" cases */
619 if (temps_are_copies(args[1], args[2])) {
620 gen_opc_buf[op_index] = op_to_movi(op);
621 tcg_opt_gen_movi(gen_args, args[0], 0);
631 /* Propagate constants through copy operations and do constant
632 folding. Constants will be substituted to arguments by register
633 allocator where needed and possible. Also detect copies. */
636 if (temps_are_copies(args[0], args[1])) {
638 gen_opc_buf[op_index] = INDEX_op_nop;
641 if (temps[args[1]].state != TCG_TEMP_CONST) {
642 tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
647 /* Source argument is constant. Rewrite the operation and
648 let movi case handle it. */
650 gen_opc_buf[op_index] = op;
651 args[1] = temps[args[1]].val;
654 tcg_opt_gen_movi(gen_args, args[0], args[1]);
661 CASE_OP_32_64(ext8s):
662 CASE_OP_32_64(ext8u):
663 CASE_OP_32_64(ext16s):
664 CASE_OP_32_64(ext16u):
665 case INDEX_op_ext32s_i64:
666 case INDEX_op_ext32u_i64:
667 if (temps[args[1]].state == TCG_TEMP_CONST) {
668 gen_opc_buf[op_index] = op_to_movi(op);
669 tmp = do_constant_folding(op, temps[args[1]].val, 0);
670 tcg_opt_gen_movi(gen_args, args[0], tmp);
693 if (temps[args[1]].state == TCG_TEMP_CONST
694 && temps[args[2]].state == TCG_TEMP_CONST) {
695 gen_opc_buf[op_index] = op_to_movi(op);
696 tmp = do_constant_folding(op, temps[args[1]].val,
698 tcg_opt_gen_movi(gen_args, args[0], tmp);
705 CASE_OP_32_64(deposit):
706 if (temps[args[1]].state == TCG_TEMP_CONST
707 && temps[args[2]].state == TCG_TEMP_CONST) {
708 gen_opc_buf[op_index] = op_to_movi(op);
709 tmp = ((1ull << args[4]) - 1);
710 tmp = (temps[args[1]].val & ~(tmp << args[3]))
711 | ((temps[args[2]].val & tmp) << args[3]);
712 tcg_opt_gen_movi(gen_args, args[0], tmp);
719 CASE_OP_32_64(setcond):
720 tmp = do_constant_folding_cond(op, args[1], args[2], args[3]);
722 gen_opc_buf[op_index] = op_to_movi(op);
723 tcg_opt_gen_movi(gen_args, args[0], tmp);
730 CASE_OP_32_64(brcond):
731 tmp = do_constant_folding_cond(op, args[0], args[1], args[2]);
734 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
735 gen_opc_buf[op_index] = INDEX_op_br;
736 gen_args[0] = args[3];
739 gen_opc_buf[op_index] = INDEX_op_nop;
746 CASE_OP_32_64(movcond):
747 tmp = do_constant_folding_cond(op, args[1], args[2], args[5]);
749 if (temps_are_copies(args[0], args[4-tmp])) {
750 gen_opc_buf[op_index] = INDEX_op_nop;
751 } else if (temps[args[4-tmp]].state == TCG_TEMP_CONST) {
752 gen_opc_buf[op_index] = op_to_movi(op);
753 tcg_opt_gen_movi(gen_args, args[0], temps[args[4-tmp]].val);
756 gen_opc_buf[op_index] = op_to_mov(op);
757 tcg_opt_gen_mov(s, gen_args, args[0], args[4-tmp]);
765 case INDEX_op_brcond2_i32:
766 /* Simplify LT/GE comparisons vs zero to a single compare
767 vs the high word of the input. */
768 if ((args[4] == TCG_COND_LT || args[4] == TCG_COND_GE)
769 && temps[args[2]].state == TCG_TEMP_CONST
770 && temps[args[3]].state == TCG_TEMP_CONST
771 && temps[args[2]].val == 0
772 && temps[args[3]].val == 0) {
773 gen_opc_buf[op_index] = INDEX_op_brcond_i32;
774 gen_args[0] = args[1];
775 gen_args[1] = args[3];
776 gen_args[2] = args[4];
777 gen_args[3] = args[5];
780 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
785 case INDEX_op_setcond2_i32:
786 /* Simplify LT/GE comparisons vs zero to a single compare
787 vs the high word of the input. */
788 if ((args[5] == TCG_COND_LT || args[5] == TCG_COND_GE)
789 && temps[args[3]].state == TCG_TEMP_CONST
790 && temps[args[4]].state == TCG_TEMP_CONST
791 && temps[args[3]].val == 0
792 && temps[args[4]].val == 0) {
793 gen_opc_buf[op_index] = INDEX_op_setcond_i32;
794 gen_args[0] = args[0];
795 gen_args[1] = args[2];
796 gen_args[2] = args[4];
797 gen_args[3] = args[5];
805 nb_call_args = (args[0] >> 16) + (args[0] & 0xffff);
806 if (!(args[nb_call_args + 1] & (TCG_CALL_CONST | TCG_CALL_PURE))) {
807 for (i = 0; i < nb_globals; i++) {
811 for (i = 0; i < (args[0] >> 16); i++) {
812 reset_temp(args[i + 1]);
814 i = nb_call_args + 3;
825 /* Default case: we know nothing about operation (or were unable
826 to compute the operation result) so no propagation is done.
827 We trash everything if the operation is the end of a basic
828 block, otherwise we only trash the output args. */
829 if (def->flags & TCG_OPF_BB_END) {
830 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
832 for (i = 0; i < def->nb_oargs; i++) {
836 for (i = 0; i < def->nb_args; i++) {
837 gen_args[i] = args[i];
839 args += def->nb_args;
840 gen_args += def->nb_args;
848 TCGArg *tcg_optimize(TCGContext *s, uint16_t *tcg_opc_ptr,
849 TCGArg *args, TCGOpDef *tcg_op_defs)
852 res = tcg_constant_folding(s, tcg_opc_ptr, args, tcg_op_defs);