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
26 #include "qemu/osdep.h"
27 #include "qemu/int128.h"
28 #include "qemu/interval-tree.h"
29 #include "tcg/tcg-op-common.h"
30 #include "tcg-internal.h"
32 #define CASE_OP_32_64(x) \
33 glue(glue(case INDEX_op_, x), _i32): \
34 glue(glue(case INDEX_op_, x), _i64)
36 #define CASE_OP_32_64_VEC(x) \
37 glue(glue(case INDEX_op_, x), _i32): \
38 glue(glue(case INDEX_op_, x), _i64): \
39 glue(glue(case INDEX_op_, x), _vec)
41 typedef struct MemCopyInfo {
42 IntervalTreeNode itree;
43 QSIMPLEQ_ENTRY (MemCopyInfo) next;
48 typedef struct TempOptInfo {
52 QSIMPLEQ_HEAD(, MemCopyInfo) mem_copy;
54 uint64_t z_mask; /* mask bit is 0 if and only if value bit is 0 */
55 uint64_t s_mask; /* a left-aligned mask of clrsb(value) bits. */
58 typedef struct OptContext {
61 TCGTempSet temps_used;
63 IntervalTreeRoot mem_copy;
64 QSIMPLEQ_HEAD(, MemCopyInfo) mem_free;
66 /* In flight values from optimization. */
67 uint64_t a_mask; /* mask bit is 0 iff value identical to first input */
68 uint64_t z_mask; /* mask bit is 0 iff value bit is 0 */
69 uint64_t s_mask; /* mask of clrsb(value) bits */
73 /* Calculate the smask for a specific value. */
74 static uint64_t smask_from_value(uint64_t value)
76 int rep = clrsb64(value);
77 return ~(~0ull >> rep);
81 * Calculate the smask for a given set of known-zeros.
82 * If there are lots of zeros on the left, we can consider the remainder
83 * an unsigned field, and thus the corresponding signed field is one bit
86 static uint64_t smask_from_zmask(uint64_t zmask)
89 * Only the 0 bits are significant for zmask, thus the msb itself
90 * must be zero, else we have no sign information.
92 int rep = clz64(zmask);
97 return ~(~0ull >> rep);
101 * Recreate a properly left-aligned smask after manipulation.
102 * Some bit-shuffling, particularly shifts and rotates, may
103 * retain sign bits on the left, but may scatter disconnected
104 * sign bits on the right. Retain only what remains to the left.
106 static uint64_t smask_from_smask(int64_t smask)
108 /* Only the 1 bits are significant for smask */
109 return smask_from_zmask(~smask);
112 static inline TempOptInfo *ts_info(TCGTemp *ts)
114 return ts->state_ptr;
117 static inline TempOptInfo *arg_info(TCGArg arg)
119 return ts_info(arg_temp(arg));
122 static inline bool ts_is_const(TCGTemp *ts)
124 return ts_info(ts)->is_const;
127 static inline bool arg_is_const(TCGArg arg)
129 return ts_is_const(arg_temp(arg));
132 static inline bool ts_is_copy(TCGTemp *ts)
134 return ts_info(ts)->next_copy != ts;
137 static TCGTemp *cmp_better_copy(TCGTemp *a, TCGTemp *b)
139 return a->kind < b->kind ? b : a;
142 /* Initialize and activate a temporary. */
143 static void init_ts_info(OptContext *ctx, TCGTemp *ts)
145 size_t idx = temp_idx(ts);
148 if (test_bit(idx, ctx->temps_used.l)) {
151 set_bit(idx, ctx->temps_used.l);
155 ti = tcg_malloc(sizeof(TempOptInfo));
161 QSIMPLEQ_INIT(&ti->mem_copy);
162 if (ts->kind == TEMP_CONST) {
165 ti->z_mask = ts->val;
166 ti->s_mask = smask_from_value(ts->val);
168 ti->is_const = false;
174 static MemCopyInfo *mem_copy_first(OptContext *ctx, intptr_t s, intptr_t l)
176 IntervalTreeNode *r = interval_tree_iter_first(&ctx->mem_copy, s, l);
177 return r ? container_of(r, MemCopyInfo, itree) : NULL;
180 static MemCopyInfo *mem_copy_next(MemCopyInfo *mem, intptr_t s, intptr_t l)
182 IntervalTreeNode *r = interval_tree_iter_next(&mem->itree, s, l);
183 return r ? container_of(r, MemCopyInfo, itree) : NULL;
186 static void remove_mem_copy(OptContext *ctx, MemCopyInfo *mc)
188 TCGTemp *ts = mc->ts;
189 TempOptInfo *ti = ts_info(ts);
191 interval_tree_remove(&mc->itree, &ctx->mem_copy);
192 QSIMPLEQ_REMOVE(&ti->mem_copy, mc, MemCopyInfo, next);
193 QSIMPLEQ_INSERT_TAIL(&ctx->mem_free, mc, next);
196 static void remove_mem_copy_in(OptContext *ctx, intptr_t s, intptr_t l)
199 MemCopyInfo *mc = mem_copy_first(ctx, s, l);
203 remove_mem_copy(ctx, mc);
207 static void remove_mem_copy_all(OptContext *ctx)
209 remove_mem_copy_in(ctx, 0, -1);
210 tcg_debug_assert(interval_tree_is_empty(&ctx->mem_copy));
213 static TCGTemp *find_better_copy(TCGTemp *ts)
217 /* If this is already readonly, we can't do better. */
218 if (temp_readonly(ts)) {
223 for (i = ts_info(ts)->next_copy; i != ts; i = ts_info(i)->next_copy) {
224 ret = cmp_better_copy(ret, i);
229 static void move_mem_copies(TCGTemp *dst_ts, TCGTemp *src_ts)
231 TempOptInfo *si = ts_info(src_ts);
232 TempOptInfo *di = ts_info(dst_ts);
235 QSIMPLEQ_FOREACH(mc, &si->mem_copy, next) {
236 tcg_debug_assert(mc->ts == src_ts);
239 QSIMPLEQ_CONCAT(&di->mem_copy, &si->mem_copy);
242 /* Reset TEMP's state, possibly removing the temp for the list of copies. */
243 static void reset_ts(OptContext *ctx, TCGTemp *ts)
245 TempOptInfo *ti = ts_info(ts);
246 TCGTemp *pts = ti->prev_copy;
247 TCGTemp *nts = ti->next_copy;
248 TempOptInfo *pi = ts_info(pts);
249 TempOptInfo *ni = ts_info(nts);
251 ni->prev_copy = ti->prev_copy;
252 pi->next_copy = ti->next_copy;
255 ti->is_const = false;
259 if (!QSIMPLEQ_EMPTY(&ti->mem_copy)) {
261 /* Last temp copy being removed, the mem copies die. */
263 QSIMPLEQ_FOREACH(mc, &ti->mem_copy, next) {
264 interval_tree_remove(&mc->itree, &ctx->mem_copy);
266 QSIMPLEQ_CONCAT(&ctx->mem_free, &ti->mem_copy);
268 move_mem_copies(find_better_copy(nts), ts);
273 static void reset_temp(OptContext *ctx, TCGArg arg)
275 reset_ts(ctx, arg_temp(arg));
278 static void record_mem_copy(OptContext *ctx, TCGType type,
279 TCGTemp *ts, intptr_t start, intptr_t last)
284 mc = QSIMPLEQ_FIRST(&ctx->mem_free);
286 QSIMPLEQ_REMOVE_HEAD(&ctx->mem_free, next);
288 mc = tcg_malloc(sizeof(*mc));
291 memset(mc, 0, sizeof(*mc));
292 mc->itree.start = start;
293 mc->itree.last = last;
295 interval_tree_insert(&mc->itree, &ctx->mem_copy);
297 ts = find_better_copy(ts);
300 QSIMPLEQ_INSERT_TAIL(&ti->mem_copy, mc, next);
303 static bool ts_are_copies(TCGTemp *ts1, TCGTemp *ts2)
311 if (!ts_is_copy(ts1) || !ts_is_copy(ts2)) {
315 for (i = ts_info(ts1)->next_copy; i != ts1; i = ts_info(i)->next_copy) {
324 static bool args_are_copies(TCGArg arg1, TCGArg arg2)
326 return ts_are_copies(arg_temp(arg1), arg_temp(arg2));
329 static TCGTemp *find_mem_copy_for(OptContext *ctx, TCGType type, intptr_t s)
333 for (mc = mem_copy_first(ctx, s, s); mc; mc = mem_copy_next(mc, s, s)) {
334 if (mc->itree.start == s && mc->type == type) {
335 return find_better_copy(mc->ts);
341 static TCGArg arg_new_constant(OptContext *ctx, uint64_t val)
343 TCGType type = ctx->type;
346 if (type == TCG_TYPE_I32) {
350 ts = tcg_constant_internal(type, val);
351 init_ts_info(ctx, ts);
356 static bool tcg_opt_gen_mov(OptContext *ctx, TCGOp *op, TCGArg dst, TCGArg src)
358 TCGTemp *dst_ts = arg_temp(dst);
359 TCGTemp *src_ts = arg_temp(src);
364 if (ts_are_copies(dst_ts, src_ts)) {
365 tcg_op_remove(ctx->tcg, op);
369 reset_ts(ctx, dst_ts);
370 di = ts_info(dst_ts);
371 si = ts_info(src_ts);
375 new_op = INDEX_op_mov_i32;
378 new_op = INDEX_op_mov_i64;
383 /* TCGOP_VECL and TCGOP_VECE remain unchanged. */
384 new_op = INDEX_op_mov_vec;
387 g_assert_not_reached();
393 di->z_mask = si->z_mask;
394 di->s_mask = si->s_mask;
396 if (src_ts->type == dst_ts->type) {
397 TempOptInfo *ni = ts_info(si->next_copy);
399 di->next_copy = si->next_copy;
400 di->prev_copy = src_ts;
401 ni->prev_copy = dst_ts;
402 si->next_copy = dst_ts;
403 di->is_const = si->is_const;
406 if (!QSIMPLEQ_EMPTY(&si->mem_copy)
407 && cmp_better_copy(src_ts, dst_ts) == dst_ts) {
408 move_mem_copies(dst_ts, src_ts);
414 static bool tcg_opt_gen_movi(OptContext *ctx, TCGOp *op,
415 TCGArg dst, uint64_t val)
417 /* Convert movi to mov with constant temp. */
418 return tcg_opt_gen_mov(ctx, op, dst, arg_new_constant(ctx, val));
421 static uint64_t do_constant_folding_2(TCGOpcode op, uint64_t x, uint64_t y)
435 CASE_OP_32_64_VEC(and):
438 CASE_OP_32_64_VEC(or):
441 CASE_OP_32_64_VEC(xor):
444 case INDEX_op_shl_i32:
445 return (uint32_t)x << (y & 31);
447 case INDEX_op_shl_i64:
448 return (uint64_t)x << (y & 63);
450 case INDEX_op_shr_i32:
451 return (uint32_t)x >> (y & 31);
453 case INDEX_op_shr_i64:
454 return (uint64_t)x >> (y & 63);
456 case INDEX_op_sar_i32:
457 return (int32_t)x >> (y & 31);
459 case INDEX_op_sar_i64:
460 return (int64_t)x >> (y & 63);
462 case INDEX_op_rotr_i32:
463 return ror32(x, y & 31);
465 case INDEX_op_rotr_i64:
466 return ror64(x, y & 63);
468 case INDEX_op_rotl_i32:
469 return rol32(x, y & 31);
471 case INDEX_op_rotl_i64:
472 return rol64(x, y & 63);
474 CASE_OP_32_64_VEC(not):
480 CASE_OP_32_64_VEC(andc):
483 CASE_OP_32_64_VEC(orc):
486 CASE_OP_32_64_VEC(eqv):
489 CASE_OP_32_64_VEC(nand):
492 CASE_OP_32_64_VEC(nor):
495 case INDEX_op_clz_i32:
496 return (uint32_t)x ? clz32(x) : y;
498 case INDEX_op_clz_i64:
499 return x ? clz64(x) : y;
501 case INDEX_op_ctz_i32:
502 return (uint32_t)x ? ctz32(x) : y;
504 case INDEX_op_ctz_i64:
505 return x ? ctz64(x) : y;
507 case INDEX_op_ctpop_i32:
510 case INDEX_op_ctpop_i64:
513 CASE_OP_32_64(ext8s):
516 CASE_OP_32_64(ext16s):
519 CASE_OP_32_64(ext8u):
522 CASE_OP_32_64(ext16u):
525 CASE_OP_32_64(bswap16):
527 return y & TCG_BSWAP_OS ? (int16_t)x : x;
529 CASE_OP_32_64(bswap32):
531 return y & TCG_BSWAP_OS ? (int32_t)x : x;
533 case INDEX_op_bswap64_i64:
536 case INDEX_op_ext_i32_i64:
537 case INDEX_op_ext32s_i64:
540 case INDEX_op_extu_i32_i64:
541 case INDEX_op_extrl_i64_i32:
542 case INDEX_op_ext32u_i64:
545 case INDEX_op_extrh_i64_i32:
546 return (uint64_t)x >> 32;
548 case INDEX_op_muluh_i32:
549 return ((uint64_t)(uint32_t)x * (uint32_t)y) >> 32;
550 case INDEX_op_mulsh_i32:
551 return ((int64_t)(int32_t)x * (int32_t)y) >> 32;
553 case INDEX_op_muluh_i64:
554 mulu64(&l64, &h64, x, y);
556 case INDEX_op_mulsh_i64:
557 muls64(&l64, &h64, x, y);
560 case INDEX_op_div_i32:
561 /* Avoid crashing on divide by zero, otherwise undefined. */
562 return (int32_t)x / ((int32_t)y ? : 1);
563 case INDEX_op_divu_i32:
564 return (uint32_t)x / ((uint32_t)y ? : 1);
565 case INDEX_op_div_i64:
566 return (int64_t)x / ((int64_t)y ? : 1);
567 case INDEX_op_divu_i64:
568 return (uint64_t)x / ((uint64_t)y ? : 1);
570 case INDEX_op_rem_i32:
571 return (int32_t)x % ((int32_t)y ? : 1);
572 case INDEX_op_remu_i32:
573 return (uint32_t)x % ((uint32_t)y ? : 1);
574 case INDEX_op_rem_i64:
575 return (int64_t)x % ((int64_t)y ? : 1);
576 case INDEX_op_remu_i64:
577 return (uint64_t)x % ((uint64_t)y ? : 1);
580 g_assert_not_reached();
584 static uint64_t do_constant_folding(TCGOpcode op, TCGType type,
585 uint64_t x, uint64_t y)
587 uint64_t res = do_constant_folding_2(op, x, y);
588 if (type == TCG_TYPE_I32) {
594 static bool do_constant_folding_cond_32(uint32_t x, uint32_t y, TCGCond c)
602 return (int32_t)x < (int32_t)y;
604 return (int32_t)x >= (int32_t)y;
606 return (int32_t)x <= (int32_t)y;
608 return (int32_t)x > (int32_t)y;
618 g_assert_not_reached();
622 static bool do_constant_folding_cond_64(uint64_t x, uint64_t y, TCGCond c)
630 return (int64_t)x < (int64_t)y;
632 return (int64_t)x >= (int64_t)y;
634 return (int64_t)x <= (int64_t)y;
636 return (int64_t)x > (int64_t)y;
646 g_assert_not_reached();
650 static bool do_constant_folding_cond_eq(TCGCond c)
666 g_assert_not_reached();
671 * Return -1 if the condition can't be simplified,
672 * and the result of the condition (0 or 1) if it can.
674 static int do_constant_folding_cond(TCGType type, TCGArg x,
677 if (arg_is_const(x) && arg_is_const(y)) {
678 uint64_t xv = arg_info(x)->val;
679 uint64_t yv = arg_info(y)->val;
683 return do_constant_folding_cond_32(xv, yv, c);
685 return do_constant_folding_cond_64(xv, yv, c);
687 /* Only scalar comparisons are optimizable */
690 } else if (args_are_copies(x, y)) {
691 return do_constant_folding_cond_eq(c);
692 } else if (arg_is_const(y) && arg_info(y)->val == 0) {
706 * Return -1 if the condition can't be simplified,
707 * and the result of the condition (0 or 1) if it can.
709 static int do_constant_folding_cond2(TCGArg *p1, TCGArg *p2, TCGCond c)
711 TCGArg al = p1[0], ah = p1[1];
712 TCGArg bl = p2[0], bh = p2[1];
714 if (arg_is_const(bl) && arg_is_const(bh)) {
715 tcg_target_ulong blv = arg_info(bl)->val;
716 tcg_target_ulong bhv = arg_info(bh)->val;
717 uint64_t b = deposit64(blv, 32, 32, bhv);
719 if (arg_is_const(al) && arg_is_const(ah)) {
720 tcg_target_ulong alv = arg_info(al)->val;
721 tcg_target_ulong ahv = arg_info(ah)->val;
722 uint64_t a = deposit64(alv, 32, 32, ahv);
723 return do_constant_folding_cond_64(a, b, c);
736 if (args_are_copies(al, bl) && args_are_copies(ah, bh)) {
737 return do_constant_folding_cond_eq(c);
744 * @dest: TCGArg of the destination argument, or NO_DEST.
745 * @p1: first paired argument
746 * @p2: second paired argument
748 * If *@p1 is a constant and *@p2 is not, swap.
749 * If *@p2 matches @dest, swap.
750 * Return true if a swap was performed.
753 #define NO_DEST temp_arg(NULL)
755 static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2)
757 TCGArg a1 = *p1, a2 = *p2;
759 sum += arg_is_const(a1);
760 sum -= arg_is_const(a2);
762 /* Prefer the constant in second argument, and then the form
763 op a, a, b, which is better handled on non-RISC hosts. */
764 if (sum > 0 || (sum == 0 && dest == a2)) {
772 static bool swap_commutative2(TCGArg *p1, TCGArg *p2)
775 sum += arg_is_const(p1[0]);
776 sum += arg_is_const(p1[1]);
777 sum -= arg_is_const(p2[0]);
778 sum -= arg_is_const(p2[1]);
781 t = p1[0], p1[0] = p2[0], p2[0] = t;
782 t = p1[1], p1[1] = p2[1], p2[1] = t;
788 static void init_arguments(OptContext *ctx, TCGOp *op, int nb_args)
790 for (int i = 0; i < nb_args; i++) {
791 TCGTemp *ts = arg_temp(op->args[i]);
792 init_ts_info(ctx, ts);
796 static void copy_propagate(OptContext *ctx, TCGOp *op,
797 int nb_oargs, int nb_iargs)
799 for (int i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
800 TCGTemp *ts = arg_temp(op->args[i]);
801 if (ts_is_copy(ts)) {
802 op->args[i] = temp_arg(find_better_copy(ts));
807 static void finish_folding(OptContext *ctx, TCGOp *op)
809 const TCGOpDef *def = &tcg_op_defs[op->opc];
813 * We only optimize extended basic blocks. If the opcode ends a BB
814 * and is not a conditional branch, reset all temp data.
816 if (def->flags & TCG_OPF_BB_END) {
818 if (!(def->flags & TCG_OPF_COND_BRANCH)) {
819 memset(&ctx->temps_used, 0, sizeof(ctx->temps_used));
820 remove_mem_copy_all(ctx);
825 nb_oargs = def->nb_oargs;
826 for (i = 0; i < nb_oargs; i++) {
827 TCGTemp *ts = arg_temp(op->args[i]);
830 * Save the corresponding known-zero/sign bits mask for the
831 * first output argument (only one supported so far).
834 ts_info(ts)->z_mask = ctx->z_mask;
835 ts_info(ts)->s_mask = ctx->s_mask;
841 * The fold_* functions return true when processing is complete,
842 * usually by folding the operation to a constant or to a copy,
843 * and calling tcg_opt_gen_{mov,movi}. They may do other things,
844 * like collect information about the value produced, for use in
845 * optimizing a subsequent operation.
847 * These first fold_* functions are all helpers, used by other
848 * folders for more specific operations.
851 static bool fold_const1(OptContext *ctx, TCGOp *op)
853 if (arg_is_const(op->args[1])) {
856 t = arg_info(op->args[1])->val;
857 t = do_constant_folding(op->opc, ctx->type, t, 0);
858 return tcg_opt_gen_movi(ctx, op, op->args[0], t);
863 static bool fold_const2(OptContext *ctx, TCGOp *op)
865 if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
866 uint64_t t1 = arg_info(op->args[1])->val;
867 uint64_t t2 = arg_info(op->args[2])->val;
869 t1 = do_constant_folding(op->opc, ctx->type, t1, t2);
870 return tcg_opt_gen_movi(ctx, op, op->args[0], t1);
875 static bool fold_commutative(OptContext *ctx, TCGOp *op)
877 swap_commutative(op->args[0], &op->args[1], &op->args[2]);
881 static bool fold_const2_commutative(OptContext *ctx, TCGOp *op)
883 swap_commutative(op->args[0], &op->args[1], &op->args[2]);
884 return fold_const2(ctx, op);
887 static bool fold_masks(OptContext *ctx, TCGOp *op)
889 uint64_t a_mask = ctx->a_mask;
890 uint64_t z_mask = ctx->z_mask;
891 uint64_t s_mask = ctx->s_mask;
894 * 32-bit ops generate 32-bit results, which for the purpose of
895 * simplifying tcg are sign-extended. Certainly that's how we
896 * represent our constants elsewhere. Note that the bits will
897 * be reset properly for a 64-bit value when encountering the
898 * type changing opcodes.
900 if (ctx->type == TCG_TYPE_I32) {
901 a_mask = (int32_t)a_mask;
902 z_mask = (int32_t)z_mask;
903 s_mask |= MAKE_64BIT_MASK(32, 32);
904 ctx->z_mask = z_mask;
905 ctx->s_mask = s_mask;
909 return tcg_opt_gen_movi(ctx, op, op->args[0], 0);
912 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
918 * Convert @op to NOT, if NOT is supported by the host.
919 * Return true f the conversion is successful, which will still
920 * indicate that the processing is complete.
922 static bool fold_not(OptContext *ctx, TCGOp *op);
923 static bool fold_to_not(OptContext *ctx, TCGOp *op, int idx)
930 not_op = INDEX_op_not_i32;
931 have_not = TCG_TARGET_HAS_not_i32;
934 not_op = INDEX_op_not_i64;
935 have_not = TCG_TARGET_HAS_not_i64;
940 not_op = INDEX_op_not_vec;
941 have_not = TCG_TARGET_HAS_not_vec;
944 g_assert_not_reached();
948 op->args[1] = op->args[idx];
949 return fold_not(ctx, op);
954 /* If the binary operation has first argument @i, fold to @i. */
955 static bool fold_ix_to_i(OptContext *ctx, TCGOp *op, uint64_t i)
957 if (arg_is_const(op->args[1]) && arg_info(op->args[1])->val == i) {
958 return tcg_opt_gen_movi(ctx, op, op->args[0], i);
963 /* If the binary operation has first argument @i, fold to NOT. */
964 static bool fold_ix_to_not(OptContext *ctx, TCGOp *op, uint64_t i)
966 if (arg_is_const(op->args[1]) && arg_info(op->args[1])->val == i) {
967 return fold_to_not(ctx, op, 2);
972 /* If the binary operation has second argument @i, fold to @i. */
973 static bool fold_xi_to_i(OptContext *ctx, TCGOp *op, uint64_t i)
975 if (arg_is_const(op->args[2]) && arg_info(op->args[2])->val == i) {
976 return tcg_opt_gen_movi(ctx, op, op->args[0], i);
981 /* If the binary operation has second argument @i, fold to identity. */
982 static bool fold_xi_to_x(OptContext *ctx, TCGOp *op, uint64_t i)
984 if (arg_is_const(op->args[2]) && arg_info(op->args[2])->val == i) {
985 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
990 /* If the binary operation has second argument @i, fold to NOT. */
991 static bool fold_xi_to_not(OptContext *ctx, TCGOp *op, uint64_t i)
993 if (arg_is_const(op->args[2]) && arg_info(op->args[2])->val == i) {
994 return fold_to_not(ctx, op, 1);
999 /* If the binary operation has both arguments equal, fold to @i. */
1000 static bool fold_xx_to_i(OptContext *ctx, TCGOp *op, uint64_t i)
1002 if (args_are_copies(op->args[1], op->args[2])) {
1003 return tcg_opt_gen_movi(ctx, op, op->args[0], i);
1008 /* If the binary operation has both arguments equal, fold to identity. */
1009 static bool fold_xx_to_x(OptContext *ctx, TCGOp *op)
1011 if (args_are_copies(op->args[1], op->args[2])) {
1012 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1018 * These outermost fold_<op> functions are sorted alphabetically.
1020 * The ordering of the transformations should be:
1021 * 1) those that produce a constant
1022 * 2) those that produce a copy
1023 * 3) those that produce information about the result value.
1026 static bool fold_add(OptContext *ctx, TCGOp *op)
1028 if (fold_const2_commutative(ctx, op) ||
1029 fold_xi_to_x(ctx, op, 0)) {
1035 /* We cannot as yet do_constant_folding with vectors. */
1036 static bool fold_add_vec(OptContext *ctx, TCGOp *op)
1038 if (fold_commutative(ctx, op) ||
1039 fold_xi_to_x(ctx, op, 0)) {
1045 static bool fold_addsub2(OptContext *ctx, TCGOp *op, bool add)
1047 bool a_const = arg_is_const(op->args[2]) && arg_is_const(op->args[3]);
1048 bool b_const = arg_is_const(op->args[4]) && arg_is_const(op->args[5]);
1050 if (a_const && b_const) {
1051 uint64_t al = arg_info(op->args[2])->val;
1052 uint64_t ah = arg_info(op->args[3])->val;
1053 uint64_t bl = arg_info(op->args[4])->val;
1054 uint64_t bh = arg_info(op->args[5])->val;
1058 if (ctx->type == TCG_TYPE_I32) {
1059 uint64_t a = deposit64(al, 32, 32, ah);
1060 uint64_t b = deposit64(bl, 32, 32, bh);
1068 al = sextract64(a, 0, 32);
1069 ah = sextract64(a, 32, 32);
1071 Int128 a = int128_make128(al, ah);
1072 Int128 b = int128_make128(bl, bh);
1075 a = int128_add(a, b);
1077 a = int128_sub(a, b);
1080 al = int128_getlo(a);
1081 ah = int128_gethi(a);
1087 /* The proper opcode is supplied by tcg_opt_gen_mov. */
1088 op2 = tcg_op_insert_before(ctx->tcg, op, 0, 2);
1090 tcg_opt_gen_movi(ctx, op, rl, al);
1091 tcg_opt_gen_movi(ctx, op2, rh, ah);
1095 /* Fold sub2 r,x,i to add2 r,x,-i */
1096 if (!add && b_const) {
1097 uint64_t bl = arg_info(op->args[4])->val;
1098 uint64_t bh = arg_info(op->args[5])->val;
1100 /* Negate the two parts without assembling and disassembling. */
1104 op->opc = (ctx->type == TCG_TYPE_I32
1105 ? INDEX_op_add2_i32 : INDEX_op_add2_i64);
1106 op->args[4] = arg_new_constant(ctx, bl);
1107 op->args[5] = arg_new_constant(ctx, bh);
1112 static bool fold_add2(OptContext *ctx, TCGOp *op)
1114 /* Note that the high and low parts may be independently swapped. */
1115 swap_commutative(op->args[0], &op->args[2], &op->args[4]);
1116 swap_commutative(op->args[1], &op->args[3], &op->args[5]);
1118 return fold_addsub2(ctx, op, true);
1121 static bool fold_and(OptContext *ctx, TCGOp *op)
1125 if (fold_const2_commutative(ctx, op) ||
1126 fold_xi_to_i(ctx, op, 0) ||
1127 fold_xi_to_x(ctx, op, -1) ||
1128 fold_xx_to_x(ctx, op)) {
1132 z1 = arg_info(op->args[1])->z_mask;
1133 z2 = arg_info(op->args[2])->z_mask;
1134 ctx->z_mask = z1 & z2;
1137 * Sign repetitions are perforce all identical, whether they are 1 or 0.
1138 * Bitwise operations preserve the relative quantity of the repetitions.
1140 ctx->s_mask = arg_info(op->args[1])->s_mask
1141 & arg_info(op->args[2])->s_mask;
1144 * Known-zeros does not imply known-ones. Therefore unless
1145 * arg2 is constant, we can't infer affected bits from it.
1147 if (arg_is_const(op->args[2])) {
1148 ctx->a_mask = z1 & ~z2;
1151 return fold_masks(ctx, op);
1154 static bool fold_andc(OptContext *ctx, TCGOp *op)
1158 if (fold_const2(ctx, op) ||
1159 fold_xx_to_i(ctx, op, 0) ||
1160 fold_xi_to_x(ctx, op, 0) ||
1161 fold_ix_to_not(ctx, op, -1)) {
1165 z1 = arg_info(op->args[1])->z_mask;
1168 * Known-zeros does not imply known-ones. Therefore unless
1169 * arg2 is constant, we can't infer anything from it.
1171 if (arg_is_const(op->args[2])) {
1172 uint64_t z2 = ~arg_info(op->args[2])->z_mask;
1173 ctx->a_mask = z1 & ~z2;
1178 ctx->s_mask = arg_info(op->args[1])->s_mask
1179 & arg_info(op->args[2])->s_mask;
1180 return fold_masks(ctx, op);
1183 static bool fold_brcond(OptContext *ctx, TCGOp *op)
1185 TCGCond cond = op->args[2];
1188 if (swap_commutative(NO_DEST, &op->args[0], &op->args[1])) {
1189 op->args[2] = cond = tcg_swap_cond(cond);
1192 i = do_constant_folding_cond(ctx->type, op->args[0], op->args[1], cond);
1194 tcg_op_remove(ctx->tcg, op);
1198 op->opc = INDEX_op_br;
1199 op->args[0] = op->args[3];
1204 static bool fold_brcond2(OptContext *ctx, TCGOp *op)
1206 TCGCond cond = op->args[4];
1207 TCGArg label = op->args[5];
1210 if (swap_commutative2(&op->args[0], &op->args[2])) {
1211 op->args[4] = cond = tcg_swap_cond(cond);
1214 i = do_constant_folding_cond2(&op->args[0], &op->args[2], cond);
1216 goto do_brcond_const;
1223 * Simplify LT/GE comparisons vs zero to a single compare
1224 * vs the high word of the input.
1226 if (arg_is_const(op->args[2]) && arg_info(op->args[2])->val == 0 &&
1227 arg_is_const(op->args[3]) && arg_info(op->args[3])->val == 0) {
1228 goto do_brcond_high;
1237 * Simplify EQ/NE comparisons where one of the pairs
1238 * can be simplified.
1240 i = do_constant_folding_cond(TCG_TYPE_I32, op->args[0],
1244 goto do_brcond_const;
1246 goto do_brcond_high;
1249 i = do_constant_folding_cond(TCG_TYPE_I32, op->args[1],
1253 goto do_brcond_const;
1255 op->opc = INDEX_op_brcond_i32;
1256 op->args[1] = op->args[2];
1258 op->args[3] = label;
1267 op->opc = INDEX_op_brcond_i32;
1268 op->args[0] = op->args[1];
1269 op->args[1] = op->args[3];
1271 op->args[3] = label;
1276 tcg_op_remove(ctx->tcg, op);
1279 op->opc = INDEX_op_br;
1280 op->args[0] = label;
1286 static bool fold_bswap(OptContext *ctx, TCGOp *op)
1288 uint64_t z_mask, s_mask, sign;
1290 if (arg_is_const(op->args[1])) {
1291 uint64_t t = arg_info(op->args[1])->val;
1293 t = do_constant_folding(op->opc, ctx->type, t, op->args[2]);
1294 return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1297 z_mask = arg_info(op->args[1])->z_mask;
1300 case INDEX_op_bswap16_i32:
1301 case INDEX_op_bswap16_i64:
1302 z_mask = bswap16(z_mask);
1305 case INDEX_op_bswap32_i32:
1306 case INDEX_op_bswap32_i64:
1307 z_mask = bswap32(z_mask);
1310 case INDEX_op_bswap64_i64:
1311 z_mask = bswap64(z_mask);
1315 g_assert_not_reached();
1317 s_mask = smask_from_zmask(z_mask);
1319 switch (op->args[2] & (TCG_BSWAP_OZ | TCG_BSWAP_OS)) {
1323 /* If the sign bit may be 1, force all the bits above to 1. */
1324 if (z_mask & sign) {
1330 /* The high bits are undefined: force all bits above the sign to 1. */
1331 z_mask |= sign << 1;
1335 ctx->z_mask = z_mask;
1336 ctx->s_mask = s_mask;
1338 return fold_masks(ctx, op);
1341 static bool fold_call(OptContext *ctx, TCGOp *op)
1343 TCGContext *s = ctx->tcg;
1344 int nb_oargs = TCGOP_CALLO(op);
1345 int nb_iargs = TCGOP_CALLI(op);
1348 init_arguments(ctx, op, nb_oargs + nb_iargs);
1349 copy_propagate(ctx, op, nb_oargs, nb_iargs);
1351 /* If the function reads or writes globals, reset temp data. */
1352 flags = tcg_call_flags(op);
1353 if (!(flags & (TCG_CALL_NO_READ_GLOBALS | TCG_CALL_NO_WRITE_GLOBALS))) {
1354 int nb_globals = s->nb_globals;
1356 for (i = 0; i < nb_globals; i++) {
1357 if (test_bit(i, ctx->temps_used.l)) {
1358 reset_ts(ctx, &ctx->tcg->temps[i]);
1363 /* If the function has side effects, reset mem data. */
1364 if (!(flags & TCG_CALL_NO_SIDE_EFFECTS)) {
1365 remove_mem_copy_all(ctx);
1368 /* Reset temp data for outputs. */
1369 for (i = 0; i < nb_oargs; i++) {
1370 reset_temp(ctx, op->args[i]);
1373 /* Stop optimizing MB across calls. */
1374 ctx->prev_mb = NULL;
1378 static bool fold_count_zeros(OptContext *ctx, TCGOp *op)
1382 if (arg_is_const(op->args[1])) {
1383 uint64_t t = arg_info(op->args[1])->val;
1386 t = do_constant_folding(op->opc, ctx->type, t, 0);
1387 return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1389 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[2]);
1392 switch (ctx->type) {
1400 g_assert_not_reached();
1402 ctx->z_mask = arg_info(op->args[2])->z_mask | z_mask;
1403 ctx->s_mask = smask_from_zmask(ctx->z_mask);
1407 static bool fold_ctpop(OptContext *ctx, TCGOp *op)
1409 if (fold_const1(ctx, op)) {
1413 switch (ctx->type) {
1415 ctx->z_mask = 32 | 31;
1418 ctx->z_mask = 64 | 63;
1421 g_assert_not_reached();
1423 ctx->s_mask = smask_from_zmask(ctx->z_mask);
1427 static bool fold_deposit(OptContext *ctx, TCGOp *op)
1431 if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
1432 uint64_t t1 = arg_info(op->args[1])->val;
1433 uint64_t t2 = arg_info(op->args[2])->val;
1435 t1 = deposit64(t1, op->args[3], op->args[4], t2);
1436 return tcg_opt_gen_movi(ctx, op, op->args[0], t1);
1439 switch (ctx->type) {
1441 and_opc = INDEX_op_and_i32;
1444 and_opc = INDEX_op_and_i64;
1447 g_assert_not_reached();
1450 /* Inserting a value into zero at offset 0. */
1451 if (arg_is_const(op->args[1])
1452 && arg_info(op->args[1])->val == 0
1453 && op->args[3] == 0) {
1454 uint64_t mask = MAKE_64BIT_MASK(0, op->args[4]);
1457 op->args[1] = op->args[2];
1458 op->args[2] = arg_new_constant(ctx, mask);
1459 ctx->z_mask = mask & arg_info(op->args[1])->z_mask;
1463 /* Inserting zero into a value. */
1464 if (arg_is_const(op->args[2])
1465 && arg_info(op->args[2])->val == 0) {
1466 uint64_t mask = deposit64(-1, op->args[3], op->args[4], 0);
1469 op->args[2] = arg_new_constant(ctx, mask);
1470 ctx->z_mask = mask & arg_info(op->args[1])->z_mask;
1474 ctx->z_mask = deposit64(arg_info(op->args[1])->z_mask,
1475 op->args[3], op->args[4],
1476 arg_info(op->args[2])->z_mask);
1480 static bool fold_divide(OptContext *ctx, TCGOp *op)
1482 if (fold_const2(ctx, op) ||
1483 fold_xi_to_x(ctx, op, 1)) {
1489 static bool fold_dup(OptContext *ctx, TCGOp *op)
1491 if (arg_is_const(op->args[1])) {
1492 uint64_t t = arg_info(op->args[1])->val;
1493 t = dup_const(TCGOP_VECE(op), t);
1494 return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1499 static bool fold_dup2(OptContext *ctx, TCGOp *op)
1501 if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
1502 uint64_t t = deposit64(arg_info(op->args[1])->val, 32, 32,
1503 arg_info(op->args[2])->val);
1504 return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1507 if (args_are_copies(op->args[1], op->args[2])) {
1508 op->opc = INDEX_op_dup_vec;
1509 TCGOP_VECE(op) = MO_32;
1514 static bool fold_eqv(OptContext *ctx, TCGOp *op)
1516 if (fold_const2_commutative(ctx, op) ||
1517 fold_xi_to_x(ctx, op, -1) ||
1518 fold_xi_to_not(ctx, op, 0)) {
1522 ctx->s_mask = arg_info(op->args[1])->s_mask
1523 & arg_info(op->args[2])->s_mask;
1527 static bool fold_extract(OptContext *ctx, TCGOp *op)
1529 uint64_t z_mask_old, z_mask;
1530 int pos = op->args[2];
1531 int len = op->args[3];
1533 if (arg_is_const(op->args[1])) {
1536 t = arg_info(op->args[1])->val;
1537 t = extract64(t, pos, len);
1538 return tcg_opt_gen_movi(ctx, op, op->args[0], t);
1541 z_mask_old = arg_info(op->args[1])->z_mask;
1542 z_mask = extract64(z_mask_old, pos, len);
1544 ctx->a_mask = z_mask_old ^ z_mask;
1546 ctx->z_mask = z_mask;
1547 ctx->s_mask = smask_from_zmask(z_mask);
1549 return fold_masks(ctx, op);
1552 static bool fold_extract2(OptContext *ctx, TCGOp *op)
1554 if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) {
1555 uint64_t v1 = arg_info(op->args[1])->val;
1556 uint64_t v2 = arg_info(op->args[2])->val;
1557 int shr = op->args[3];
1559 if (op->opc == INDEX_op_extract2_i64) {
1563 v1 = (uint32_t)v1 >> shr;
1564 v2 = (uint64_t)((int32_t)v2 << (32 - shr));
1566 return tcg_opt_gen_movi(ctx, op, op->args[0], v1 | v2);
1571 static bool fold_exts(OptContext *ctx, TCGOp *op)
1573 uint64_t s_mask_old, s_mask, z_mask, sign;
1574 bool type_change = false;
1576 if (fold_const1(ctx, op)) {
1580 z_mask = arg_info(op->args[1])->z_mask;
1581 s_mask = arg_info(op->args[1])->s_mask;
1582 s_mask_old = s_mask;
1585 CASE_OP_32_64(ext8s):
1587 z_mask = (uint8_t)z_mask;
1589 CASE_OP_32_64(ext16s):
1591 z_mask = (uint16_t)z_mask;
1593 case INDEX_op_ext_i32_i64:
1596 case INDEX_op_ext32s_i64:
1598 z_mask = (uint32_t)z_mask;
1601 g_assert_not_reached();
1604 if (z_mask & sign) {
1607 s_mask |= sign << 1;
1609 ctx->z_mask = z_mask;
1610 ctx->s_mask = s_mask;
1612 ctx->a_mask = s_mask & ~s_mask_old;
1615 return fold_masks(ctx, op);
1618 static bool fold_extu(OptContext *ctx, TCGOp *op)
1620 uint64_t z_mask_old, z_mask;
1621 bool type_change = false;
1623 if (fold_const1(ctx, op)) {
1627 z_mask_old = z_mask = arg_info(op->args[1])->z_mask;
1630 CASE_OP_32_64(ext8u):
1631 z_mask = (uint8_t)z_mask;
1633 CASE_OP_32_64(ext16u):
1634 z_mask = (uint16_t)z_mask;
1636 case INDEX_op_extrl_i64_i32:
1637 case INDEX_op_extu_i32_i64:
1640 case INDEX_op_ext32u_i64:
1641 z_mask = (uint32_t)z_mask;
1643 case INDEX_op_extrh_i64_i32:
1648 g_assert_not_reached();
1651 ctx->z_mask = z_mask;
1652 ctx->s_mask = smask_from_zmask(z_mask);
1654 ctx->a_mask = z_mask_old ^ z_mask;
1656 return fold_masks(ctx, op);
1659 static bool fold_mb(OptContext *ctx, TCGOp *op)
1661 /* Eliminate duplicate and redundant fence instructions. */
1664 * Merge two barriers of the same type into one,
1665 * or a weaker barrier into a stronger one,
1666 * or two weaker barriers into a stronger one.
1667 * mb X; mb Y => mb X|Y
1668 * mb; strl => mb; st
1669 * ldaq; mb => ld; mb
1670 * ldaq; strl => ld; mb; st
1671 * Other combinations are also merged into a strong
1672 * barrier. This is stricter than specified but for
1673 * the purposes of TCG is better than not optimizing.
1675 ctx->prev_mb->args[0] |= op->args[0];
1676 tcg_op_remove(ctx->tcg, op);
1683 static bool fold_mov(OptContext *ctx, TCGOp *op)
1685 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]);
1688 static bool fold_movcond(OptContext *ctx, TCGOp *op)
1690 TCGCond cond = op->args[5];
1693 if (swap_commutative(NO_DEST, &op->args[1], &op->args[2])) {
1694 op->args[5] = cond = tcg_swap_cond(cond);
1697 * Canonicalize the "false" input reg to match the destination reg so
1698 * that the tcg backend can implement a "move if true" operation.
1700 if (swap_commutative(op->args[0], &op->args[4], &op->args[3])) {
1701 op->args[5] = cond = tcg_invert_cond(cond);
1704 i = do_constant_folding_cond(ctx->type, op->args[1], op->args[2], cond);
1706 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[4 - i]);
1709 ctx->z_mask = arg_info(op->args[3])->z_mask
1710 | arg_info(op->args[4])->z_mask;
1711 ctx->s_mask = arg_info(op->args[3])->s_mask
1712 & arg_info(op->args[4])->s_mask;
1714 if (arg_is_const(op->args[3]) && arg_is_const(op->args[4])) {
1715 uint64_t tv = arg_info(op->args[3])->val;
1716 uint64_t fv = arg_info(op->args[4])->val;
1717 TCGOpcode opc, negopc = 0;
1719 switch (ctx->type) {
1721 opc = INDEX_op_setcond_i32;
1722 if (TCG_TARGET_HAS_negsetcond_i32) {
1723 negopc = INDEX_op_negsetcond_i32;
1729 opc = INDEX_op_setcond_i64;
1730 if (TCG_TARGET_HAS_negsetcond_i64) {
1731 negopc = INDEX_op_negsetcond_i64;
1735 g_assert_not_reached();
1738 if (tv == 1 && fv == 0) {
1741 } else if (fv == 1 && tv == 0) {
1743 op->args[3] = tcg_invert_cond(cond);
1744 } else if (negopc) {
1745 if (tv == -1 && fv == 0) {
1748 } else if (fv == -1 && tv == 0) {
1750 op->args[3] = tcg_invert_cond(cond);
1757 static bool fold_mul(OptContext *ctx, TCGOp *op)
1759 if (fold_const2(ctx, op) ||
1760 fold_xi_to_i(ctx, op, 0) ||
1761 fold_xi_to_x(ctx, op, 1)) {
1767 static bool fold_mul_highpart(OptContext *ctx, TCGOp *op)
1769 if (fold_const2_commutative(ctx, op) ||
1770 fold_xi_to_i(ctx, op, 0)) {
1776 static bool fold_multiply2(OptContext *ctx, TCGOp *op)
1778 swap_commutative(op->args[0], &op->args[2], &op->args[3]);
1780 if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) {
1781 uint64_t a = arg_info(op->args[2])->val;
1782 uint64_t b = arg_info(op->args[3])->val;
1788 case INDEX_op_mulu2_i32:
1789 l = (uint64_t)(uint32_t)a * (uint32_t)b;
1790 h = (int32_t)(l >> 32);
1793 case INDEX_op_muls2_i32:
1794 l = (int64_t)(int32_t)a * (int32_t)b;
1798 case INDEX_op_mulu2_i64:
1799 mulu64(&l, &h, a, b);
1801 case INDEX_op_muls2_i64:
1802 muls64(&l, &h, a, b);
1805 g_assert_not_reached();
1811 /* The proper opcode is supplied by tcg_opt_gen_mov. */
1812 op2 = tcg_op_insert_before(ctx->tcg, op, 0, 2);
1814 tcg_opt_gen_movi(ctx, op, rl, l);
1815 tcg_opt_gen_movi(ctx, op2, rh, h);
1821 static bool fold_nand(OptContext *ctx, TCGOp *op)
1823 if (fold_const2_commutative(ctx, op) ||
1824 fold_xi_to_not(ctx, op, -1)) {
1828 ctx->s_mask = arg_info(op->args[1])->s_mask
1829 & arg_info(op->args[2])->s_mask;
1833 static bool fold_neg(OptContext *ctx, TCGOp *op)
1837 if (fold_const1(ctx, op)) {
1841 /* Set to 1 all bits to the left of the rightmost. */
1842 z_mask = arg_info(op->args[1])->z_mask;
1843 ctx->z_mask = -(z_mask & -z_mask);
1846 * Because of fold_sub_to_neg, we want to always return true,
1847 * via finish_folding.
1849 finish_folding(ctx, op);
1853 static bool fold_nor(OptContext *ctx, TCGOp *op)
1855 if (fold_const2_commutative(ctx, op) ||
1856 fold_xi_to_not(ctx, op, 0)) {
1860 ctx->s_mask = arg_info(op->args[1])->s_mask
1861 & arg_info(op->args[2])->s_mask;
1865 static bool fold_not(OptContext *ctx, TCGOp *op)
1867 if (fold_const1(ctx, op)) {
1871 ctx->s_mask = arg_info(op->args[1])->s_mask;
1873 /* Because of fold_to_not, we want to always return true, via finish. */
1874 finish_folding(ctx, op);
1878 static bool fold_or(OptContext *ctx, TCGOp *op)
1880 if (fold_const2_commutative(ctx, op) ||
1881 fold_xi_to_x(ctx, op, 0) ||
1882 fold_xx_to_x(ctx, op)) {
1886 ctx->z_mask = arg_info(op->args[1])->z_mask
1887 | arg_info(op->args[2])->z_mask;
1888 ctx->s_mask = arg_info(op->args[1])->s_mask
1889 & arg_info(op->args[2])->s_mask;
1890 return fold_masks(ctx, op);
1893 static bool fold_orc(OptContext *ctx, TCGOp *op)
1895 if (fold_const2(ctx, op) ||
1896 fold_xx_to_i(ctx, op, -1) ||
1897 fold_xi_to_x(ctx, op, -1) ||
1898 fold_ix_to_not(ctx, op, 0)) {
1902 ctx->s_mask = arg_info(op->args[1])->s_mask
1903 & arg_info(op->args[2])->s_mask;
1907 static bool fold_qemu_ld(OptContext *ctx, TCGOp *op)
1909 const TCGOpDef *def = &tcg_op_defs[op->opc];
1910 MemOpIdx oi = op->args[def->nb_oargs + def->nb_iargs];
1911 MemOp mop = get_memop(oi);
1912 int width = 8 * memop_size(mop);
1915 ctx->s_mask = MAKE_64BIT_MASK(width, 64 - width);
1916 if (!(mop & MO_SIGN)) {
1917 ctx->z_mask = MAKE_64BIT_MASK(0, width);
1922 /* Opcodes that touch guest memory stop the mb optimization. */
1923 ctx->prev_mb = NULL;
1927 static bool fold_qemu_st(OptContext *ctx, TCGOp *op)
1929 /* Opcodes that touch guest memory stop the mb optimization. */
1930 ctx->prev_mb = NULL;
1934 static bool fold_remainder(OptContext *ctx, TCGOp *op)
1936 if (fold_const2(ctx, op) ||
1937 fold_xx_to_i(ctx, op, 0)) {
1943 static bool fold_setcond(OptContext *ctx, TCGOp *op)
1945 TCGCond cond = op->args[3];
1948 if (swap_commutative(op->args[0], &op->args[1], &op->args[2])) {
1949 op->args[3] = cond = tcg_swap_cond(cond);
1952 i = do_constant_folding_cond(ctx->type, op->args[1], op->args[2], cond);
1954 return tcg_opt_gen_movi(ctx, op, op->args[0], i);
1958 ctx->s_mask = smask_from_zmask(1);
1962 static bool fold_negsetcond(OptContext *ctx, TCGOp *op)
1964 TCGCond cond = op->args[3];
1967 if (swap_commutative(op->args[0], &op->args[1], &op->args[2])) {
1968 op->args[3] = cond = tcg_swap_cond(cond);
1971 i = do_constant_folding_cond(ctx->type, op->args[1], op->args[2], cond);
1973 return tcg_opt_gen_movi(ctx, op, op->args[0], -i);
1976 /* Value is {0,-1} so all bits are repetitions of the sign. */
1982 static bool fold_setcond2(OptContext *ctx, TCGOp *op)
1984 TCGCond cond = op->args[5];
1987 if (swap_commutative2(&op->args[1], &op->args[3])) {
1988 op->args[5] = cond = tcg_swap_cond(cond);
1991 i = do_constant_folding_cond2(&op->args[1], &op->args[3], cond);
1993 goto do_setcond_const;
2000 * Simplify LT/GE comparisons vs zero to a single compare
2001 * vs the high word of the input.
2003 if (arg_is_const(op->args[3]) && arg_info(op->args[3])->val == 0 &&
2004 arg_is_const(op->args[4]) && arg_info(op->args[4])->val == 0) {
2005 goto do_setcond_high;
2014 * Simplify EQ/NE comparisons where one of the pairs
2015 * can be simplified.
2017 i = do_constant_folding_cond(TCG_TYPE_I32, op->args[1],
2021 goto do_setcond_const;
2023 goto do_setcond_high;
2026 i = do_constant_folding_cond(TCG_TYPE_I32, op->args[2],
2030 goto do_setcond_const;
2032 op->args[2] = op->args[3];
2034 op->opc = INDEX_op_setcond_i32;
2043 op->args[1] = op->args[2];
2044 op->args[2] = op->args[4];
2046 op->opc = INDEX_op_setcond_i32;
2051 ctx->s_mask = smask_from_zmask(1);
2055 return tcg_opt_gen_movi(ctx, op, op->args[0], i);
2058 static bool fold_sextract(OptContext *ctx, TCGOp *op)
2060 uint64_t z_mask, s_mask, s_mask_old;
2061 int pos = op->args[2];
2062 int len = op->args[3];
2064 if (arg_is_const(op->args[1])) {
2067 t = arg_info(op->args[1])->val;
2068 t = sextract64(t, pos, len);
2069 return tcg_opt_gen_movi(ctx, op, op->args[0], t);
2072 z_mask = arg_info(op->args[1])->z_mask;
2073 z_mask = sextract64(z_mask, pos, len);
2074 ctx->z_mask = z_mask;
2076 s_mask_old = arg_info(op->args[1])->s_mask;
2077 s_mask = sextract64(s_mask_old, pos, len);
2078 s_mask |= MAKE_64BIT_MASK(len, 64 - len);
2079 ctx->s_mask = s_mask;
2082 ctx->a_mask = s_mask & ~s_mask_old;
2085 return fold_masks(ctx, op);
2088 static bool fold_shift(OptContext *ctx, TCGOp *op)
2090 uint64_t s_mask, z_mask, sign;
2092 if (fold_const2(ctx, op) ||
2093 fold_ix_to_i(ctx, op, 0) ||
2094 fold_xi_to_x(ctx, op, 0)) {
2098 s_mask = arg_info(op->args[1])->s_mask;
2099 z_mask = arg_info(op->args[1])->z_mask;
2101 if (arg_is_const(op->args[2])) {
2102 int sh = arg_info(op->args[2])->val;
2104 ctx->z_mask = do_constant_folding(op->opc, ctx->type, z_mask, sh);
2106 s_mask = do_constant_folding(op->opc, ctx->type, s_mask, sh);
2107 ctx->s_mask = smask_from_smask(s_mask);
2109 return fold_masks(ctx, op);
2115 * Arithmetic right shift will not reduce the number of
2116 * input sign repetitions.
2118 ctx->s_mask = s_mask;
2122 * If the sign bit is known zero, then logical right shift
2123 * will not reduced the number of input sign repetitions.
2125 sign = (s_mask & -s_mask) >> 1;
2126 if (!(z_mask & sign)) {
2127 ctx->s_mask = s_mask;
2137 static bool fold_sub_to_neg(OptContext *ctx, TCGOp *op)
2142 if (!arg_is_const(op->args[1]) || arg_info(op->args[1])->val != 0) {
2146 switch (ctx->type) {
2148 neg_op = INDEX_op_neg_i32;
2152 neg_op = INDEX_op_neg_i64;
2158 neg_op = INDEX_op_neg_vec;
2159 have_neg = (TCG_TARGET_HAS_neg_vec &&
2160 tcg_can_emit_vec_op(neg_op, ctx->type, TCGOP_VECE(op)) > 0);
2163 g_assert_not_reached();
2167 op->args[1] = op->args[2];
2168 return fold_neg(ctx, op);
2173 /* We cannot as yet do_constant_folding with vectors. */
2174 static bool fold_sub_vec(OptContext *ctx, TCGOp *op)
2176 if (fold_xx_to_i(ctx, op, 0) ||
2177 fold_xi_to_x(ctx, op, 0) ||
2178 fold_sub_to_neg(ctx, op)) {
2184 static bool fold_sub(OptContext *ctx, TCGOp *op)
2186 if (fold_const2(ctx, op) || fold_sub_vec(ctx, op)) {
2190 /* Fold sub r,x,i to add r,x,-i */
2191 if (arg_is_const(op->args[2])) {
2192 uint64_t val = arg_info(op->args[2])->val;
2194 op->opc = (ctx->type == TCG_TYPE_I32
2195 ? INDEX_op_add_i32 : INDEX_op_add_i64);
2196 op->args[2] = arg_new_constant(ctx, -val);
2201 static bool fold_sub2(OptContext *ctx, TCGOp *op)
2203 return fold_addsub2(ctx, op, false);
2206 static bool fold_tcg_ld(OptContext *ctx, TCGOp *op)
2208 /* We can't do any folding with a load, but we can record bits. */
2210 CASE_OP_32_64(ld8s):
2211 ctx->s_mask = MAKE_64BIT_MASK(8, 56);
2213 CASE_OP_32_64(ld8u):
2214 ctx->z_mask = MAKE_64BIT_MASK(0, 8);
2215 ctx->s_mask = MAKE_64BIT_MASK(9, 55);
2217 CASE_OP_32_64(ld16s):
2218 ctx->s_mask = MAKE_64BIT_MASK(16, 48);
2220 CASE_OP_32_64(ld16u):
2221 ctx->z_mask = MAKE_64BIT_MASK(0, 16);
2222 ctx->s_mask = MAKE_64BIT_MASK(17, 47);
2224 case INDEX_op_ld32s_i64:
2225 ctx->s_mask = MAKE_64BIT_MASK(32, 32);
2227 case INDEX_op_ld32u_i64:
2228 ctx->z_mask = MAKE_64BIT_MASK(0, 32);
2229 ctx->s_mask = MAKE_64BIT_MASK(33, 31);
2232 g_assert_not_reached();
2237 static bool fold_tcg_ld_memcopy(OptContext *ctx, TCGOp *op)
2243 if (op->args[1] != tcgv_ptr_arg(tcg_env)) {
2249 dst = arg_temp(op->args[0]);
2250 src = find_mem_copy_for(ctx, type, ofs);
2251 if (src && src->base_type == type) {
2252 return tcg_opt_gen_mov(ctx, op, temp_arg(dst), temp_arg(src));
2256 record_mem_copy(ctx, type, dst, ofs, ofs + tcg_type_size(type) - 1);
2260 static bool fold_tcg_st(OptContext *ctx, TCGOp *op)
2262 intptr_t ofs = op->args[2];
2265 if (op->args[1] != tcgv_ptr_arg(tcg_env)) {
2266 remove_mem_copy_all(ctx);
2274 CASE_OP_32_64(st16):
2277 case INDEX_op_st32_i64:
2278 case INDEX_op_st_i32:
2281 case INDEX_op_st_i64:
2284 case INDEX_op_st_vec:
2285 lm1 = tcg_type_size(ctx->type) - 1;
2288 g_assert_not_reached();
2290 remove_mem_copy_in(ctx, ofs, ofs + lm1);
2294 static bool fold_tcg_st_memcopy(OptContext *ctx, TCGOp *op)
2300 if (op->args[1] != tcgv_ptr_arg(tcg_env)) {
2301 fold_tcg_st(ctx, op);
2305 src = arg_temp(op->args[0]);
2310 * Eliminate duplicate stores of a constant.
2311 * This happens frequently when the target ISA zero-extends.
2313 if (ts_is_const(src)) {
2314 TCGTemp *prev = find_mem_copy_for(ctx, type, ofs);
2316 tcg_op_remove(ctx->tcg, op);
2321 last = ofs + tcg_type_size(type) - 1;
2322 remove_mem_copy_in(ctx, ofs, last);
2323 record_mem_copy(ctx, type, src, ofs, last);
2327 static bool fold_xor(OptContext *ctx, TCGOp *op)
2329 if (fold_const2_commutative(ctx, op) ||
2330 fold_xx_to_i(ctx, op, 0) ||
2331 fold_xi_to_x(ctx, op, 0) ||
2332 fold_xi_to_not(ctx, op, -1)) {
2336 ctx->z_mask = arg_info(op->args[1])->z_mask
2337 | arg_info(op->args[2])->z_mask;
2338 ctx->s_mask = arg_info(op->args[1])->s_mask
2339 & arg_info(op->args[2])->s_mask;
2340 return fold_masks(ctx, op);
2343 /* Propagate constants and copies, fold constant expressions. */
2344 void tcg_optimize(TCGContext *s)
2347 TCGOp *op, *op_next;
2348 OptContext ctx = { .tcg = s };
2350 QSIMPLEQ_INIT(&ctx.mem_free);
2352 /* Array VALS has an element for each temp.
2353 If this temp holds a constant then its value is kept in VALS' element.
2354 If this temp is a copy of other ones then the other copies are
2355 available through the doubly linked circular list. */
2357 nb_temps = s->nb_temps;
2358 for (i = 0; i < nb_temps; ++i) {
2359 s->temps[i].state_ptr = NULL;
2362 QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) {
2363 TCGOpcode opc = op->opc;
2364 const TCGOpDef *def;
2367 /* Calls are special. */
2368 if (opc == INDEX_op_call) {
2369 fold_call(&ctx, op);
2373 def = &tcg_op_defs[opc];
2374 init_arguments(&ctx, op, def->nb_oargs + def->nb_iargs);
2375 copy_propagate(&ctx, op, def->nb_oargs, def->nb_iargs);
2377 /* Pre-compute the type of the operation. */
2378 if (def->flags & TCG_OPF_VECTOR) {
2379 ctx.type = TCG_TYPE_V64 + TCGOP_VECL(op);
2380 } else if (def->flags & TCG_OPF_64BIT) {
2381 ctx.type = TCG_TYPE_I64;
2383 ctx.type = TCG_TYPE_I32;
2386 /* Assume all bits affected, no bits known zero, no sign reps. */
2392 * Process each opcode.
2393 * Sorted alphabetically by opcode as much as possible.
2397 done = fold_add(&ctx, op);
2399 case INDEX_op_add_vec:
2400 done = fold_add_vec(&ctx, op);
2402 CASE_OP_32_64(add2):
2403 done = fold_add2(&ctx, op);
2405 CASE_OP_32_64_VEC(and):
2406 done = fold_and(&ctx, op);
2408 CASE_OP_32_64_VEC(andc):
2409 done = fold_andc(&ctx, op);
2411 CASE_OP_32_64(brcond):
2412 done = fold_brcond(&ctx, op);
2414 case INDEX_op_brcond2_i32:
2415 done = fold_brcond2(&ctx, op);
2417 CASE_OP_32_64(bswap16):
2418 CASE_OP_32_64(bswap32):
2419 case INDEX_op_bswap64_i64:
2420 done = fold_bswap(&ctx, op);
2424 done = fold_count_zeros(&ctx, op);
2426 CASE_OP_32_64(ctpop):
2427 done = fold_ctpop(&ctx, op);
2429 CASE_OP_32_64(deposit):
2430 done = fold_deposit(&ctx, op);
2433 CASE_OP_32_64(divu):
2434 done = fold_divide(&ctx, op);
2436 case INDEX_op_dup_vec:
2437 done = fold_dup(&ctx, op);
2439 case INDEX_op_dup2_vec:
2440 done = fold_dup2(&ctx, op);
2442 CASE_OP_32_64_VEC(eqv):
2443 done = fold_eqv(&ctx, op);
2445 CASE_OP_32_64(extract):
2446 done = fold_extract(&ctx, op);
2448 CASE_OP_32_64(extract2):
2449 done = fold_extract2(&ctx, op);
2451 CASE_OP_32_64(ext8s):
2452 CASE_OP_32_64(ext16s):
2453 case INDEX_op_ext32s_i64:
2454 case INDEX_op_ext_i32_i64:
2455 done = fold_exts(&ctx, op);
2457 CASE_OP_32_64(ext8u):
2458 CASE_OP_32_64(ext16u):
2459 case INDEX_op_ext32u_i64:
2460 case INDEX_op_extu_i32_i64:
2461 case INDEX_op_extrl_i64_i32:
2462 case INDEX_op_extrh_i64_i32:
2463 done = fold_extu(&ctx, op);
2465 CASE_OP_32_64(ld8s):
2466 CASE_OP_32_64(ld8u):
2467 CASE_OP_32_64(ld16s):
2468 CASE_OP_32_64(ld16u):
2469 case INDEX_op_ld32s_i64:
2470 case INDEX_op_ld32u_i64:
2471 done = fold_tcg_ld(&ctx, op);
2473 case INDEX_op_ld_i32:
2474 case INDEX_op_ld_i64:
2475 case INDEX_op_ld_vec:
2476 done = fold_tcg_ld_memcopy(&ctx, op);
2479 CASE_OP_32_64(st16):
2480 case INDEX_op_st32_i64:
2481 done = fold_tcg_st(&ctx, op);
2483 case INDEX_op_st_i32:
2484 case INDEX_op_st_i64:
2485 case INDEX_op_st_vec:
2486 done = fold_tcg_st_memcopy(&ctx, op);
2489 done = fold_mb(&ctx, op);
2491 CASE_OP_32_64_VEC(mov):
2492 done = fold_mov(&ctx, op);
2494 CASE_OP_32_64(movcond):
2495 done = fold_movcond(&ctx, op);
2498 done = fold_mul(&ctx, op);
2500 CASE_OP_32_64(mulsh):
2501 CASE_OP_32_64(muluh):
2502 done = fold_mul_highpart(&ctx, op);
2504 CASE_OP_32_64(muls2):
2505 CASE_OP_32_64(mulu2):
2506 done = fold_multiply2(&ctx, op);
2508 CASE_OP_32_64_VEC(nand):
2509 done = fold_nand(&ctx, op);
2512 done = fold_neg(&ctx, op);
2514 CASE_OP_32_64_VEC(nor):
2515 done = fold_nor(&ctx, op);
2517 CASE_OP_32_64_VEC(not):
2518 done = fold_not(&ctx, op);
2520 CASE_OP_32_64_VEC(or):
2521 done = fold_or(&ctx, op);
2523 CASE_OP_32_64_VEC(orc):
2524 done = fold_orc(&ctx, op);
2526 case INDEX_op_qemu_ld_a32_i32:
2527 case INDEX_op_qemu_ld_a64_i32:
2528 case INDEX_op_qemu_ld_a32_i64:
2529 case INDEX_op_qemu_ld_a64_i64:
2530 case INDEX_op_qemu_ld_a32_i128:
2531 case INDEX_op_qemu_ld_a64_i128:
2532 done = fold_qemu_ld(&ctx, op);
2534 case INDEX_op_qemu_st8_a32_i32:
2535 case INDEX_op_qemu_st8_a64_i32:
2536 case INDEX_op_qemu_st_a32_i32:
2537 case INDEX_op_qemu_st_a64_i32:
2538 case INDEX_op_qemu_st_a32_i64:
2539 case INDEX_op_qemu_st_a64_i64:
2540 case INDEX_op_qemu_st_a32_i128:
2541 case INDEX_op_qemu_st_a64_i128:
2542 done = fold_qemu_st(&ctx, op);
2545 CASE_OP_32_64(remu):
2546 done = fold_remainder(&ctx, op);
2548 CASE_OP_32_64(rotl):
2549 CASE_OP_32_64(rotr):
2553 done = fold_shift(&ctx, op);
2555 CASE_OP_32_64(setcond):
2556 done = fold_setcond(&ctx, op);
2558 CASE_OP_32_64(negsetcond):
2559 done = fold_negsetcond(&ctx, op);
2561 case INDEX_op_setcond2_i32:
2562 done = fold_setcond2(&ctx, op);
2564 CASE_OP_32_64(sextract):
2565 done = fold_sextract(&ctx, op);
2568 done = fold_sub(&ctx, op);
2570 case INDEX_op_sub_vec:
2571 done = fold_sub_vec(&ctx, op);
2573 CASE_OP_32_64(sub2):
2574 done = fold_sub2(&ctx, op);
2576 CASE_OP_32_64_VEC(xor):
2577 done = fold_xor(&ctx, op);
2584 finish_folding(&ctx, op);