2 * Simplify - do instruction simplification before CSE
4 * Copyright (C) 2004 Linus Torvalds
10 #include "expression.h"
11 #include "linearize.h"
15 /* Find the trivial parent for a phi-source */
16 static struct basic_block
*phi_parent(struct basic_block
*source
, pseudo_t pseudo
)
18 /* Can't go upwards if the pseudo is defined in the bb it came from.. */
19 if (pseudo
->type
== PSEUDO_REG
) {
20 struct instruction
*def
= pseudo
->def
;
21 if (def
->bb
== source
)
24 if (bb_list_size(source
->children
) != 1 || bb_list_size(source
->parents
) != 1)
26 return first_basic_block(source
->parents
);
29 static void clear_phi(struct instruction
*insn
)
34 FOR_EACH_PTR(insn
->phi_list
, phi
) {
35 *THIS_ADDRESS(phi
) = VOID
;
36 } END_FOR_EACH_PTR(phi
);
39 static int if_convert_phi(struct instruction
*insn
)
42 struct basic_block
*parents
[3];
43 struct basic_block
*bb
, *bb1
, *bb2
, *source
;
44 struct instruction
*br
;
48 if (linearize_ptr_list((struct ptr_list
*)insn
->phi_list
, (void **)array
, 3) != 2)
50 if (linearize_ptr_list((struct ptr_list
*)bb
->parents
, (void **)parents
, 3) != 2)
52 p1
= array
[0]->def
->src1
;
53 bb1
= array
[0]->def
->bb
;
54 p2
= array
[1]->def
->src1
;
55 bb2
= array
[1]->def
->bb
;
57 /* Only try the simple "direct parents" case */
58 if ((bb1
!= parents
[0] || bb2
!= parents
[1]) &&
59 (bb1
!= parents
[1] || bb2
!= parents
[0]))
63 * See if we can find a common source for this..
65 source
= phi_parent(bb1
, p1
);
66 if (source
!= phi_parent(bb2
, p2
))
70 * Cool. We now know that 'source' is the exclusive
71 * parent of both phi-nodes, so the exit at the
72 * end of it fully determines which one it is, and
73 * we can turn it into a select.
75 * HOWEVER, right now we only handle regular
76 * conditional branches. No multijumps or computed
77 * stuff. Verify that here.
79 br
= last_instruction(source
->insns
);
80 if (!br
|| br
->opcode
!= OP_BR
)
87 * We're in business. Match up true/false with p1/p2.
89 if (br
->bb_true
== bb2
|| br
->bb_false
== bb1
) {
96 * OK, we can now replace that last
103 * select pseudo, p1, p2
106 * and remove the phi-node. If it then
107 * turns out that 'a' or 'b' is entirely
108 * empty (common case), and now no longer
109 * a phi-source, we'll be able to simplify
110 * the conditional branch too.
112 insert_select(source
, br
, insn
, p1
, p2
);
117 static int clean_up_phi(struct instruction
*insn
)
120 struct instruction
*last
;
125 FOR_EACH_PTR(insn
->phi_list
, phi
) {
126 struct instruction
*def
;
130 if (def
->src1
== VOID
|| !def
->bb
)
133 if (last
->src1
!= def
->src1
)
138 } END_FOR_EACH_PTR(phi
);
141 pseudo_t pseudo
= last
? last
->src1
: VOID
;
142 convert_instruction_target(insn
, pseudo
);
147 return if_convert_phi(insn
);
150 static int delete_pseudo_user_list_entry(struct pseudo_user_list
**list
, pseudo_t
*entry
, int count
)
152 struct pseudo_user
*pu
;
154 FOR_EACH_PTR(*list
, pu
) {
155 if (pu
->userp
== entry
) {
156 DELETE_CURRENT_PTR(pu
);
160 } END_FOR_EACH_PTR(pu
);
163 pack_ptr_list((struct ptr_list
**)list
);
167 static inline void remove_usage(pseudo_t p
, pseudo_t
*usep
)
169 if (has_use_list(p
)) {
170 delete_pseudo_user_list_entry(&p
->users
, usep
, 1);
172 kill_instruction(p
->def
);
176 void kill_use(pseudo_t
*usep
)
181 remove_usage(p
, usep
);
185 void kill_instruction(struct instruction
*insn
)
187 if (!insn
|| !insn
->bb
)
190 switch (insn
->opcode
) {
193 kill_use(&insn
->src3
);
196 case OP_BINARY
... OP_BINCMP_END
:
197 kill_use(&insn
->src2
);
205 case OP_NOT
: case OP_NEG
:
206 kill_use(&insn
->src1
);
214 repeat_phase
|= REPEAT_SYMBOL_CLEANUP
;
218 if (!insn
->bb_true
|| !insn
->bb_false
)
222 case OP_COMPUTEDGOTO
:
223 kill_use(&insn
->cond
);
233 repeat_phase
|= REPEAT_CSE
;
238 * Kill trivially dead instructions
240 static int dead_insn(struct instruction
*insn
, pseudo_t
*src1
, pseudo_t
*src2
, pseudo_t
*src3
)
242 struct pseudo_user
*pu
;
243 FOR_EACH_PTR(insn
->target
->users
, pu
) {
244 if (*pu
->userp
!= VOID
)
246 } END_FOR_EACH_PTR(pu
);
255 static inline int constant(pseudo_t pseudo
)
257 return pseudo
->type
== PSEUDO_VAL
;
260 static int replace_with_pseudo(struct instruction
*insn
, pseudo_t pseudo
)
262 convert_instruction_target(insn
, pseudo
);
264 switch (insn
->opcode
) {
267 kill_use(&insn
->src3
);
268 case OP_BINARY
... OP_BINCMP_END
:
269 kill_use(&insn
->src2
);
277 kill_use(&insn
->src1
);
287 static unsigned int value_size(long long value
)
302 * Try to determine the maximum size of bits in a pseudo.
304 * Right now this only follow casts and constant values, but we
305 * could look at things like logical 'and' instructions etc.
307 static unsigned int operand_size(struct instruction
*insn
, pseudo_t pseudo
)
309 unsigned int size
= insn
->size
;
311 if (pseudo
->type
== PSEUDO_REG
) {
312 struct instruction
*src
= pseudo
->def
;
313 if (src
&& src
->opcode
== OP_CAST
&& src
->orig_type
) {
314 unsigned int orig_size
= src
->orig_type
->bit_size
;
315 if (orig_size
< size
)
319 if (pseudo
->type
== PSEUDO_VAL
) {
320 unsigned int orig_size
= value_size(pseudo
->value
);
321 if (orig_size
< size
)
327 static int simplify_asr(struct instruction
*insn
, pseudo_t pseudo
, long long value
)
329 unsigned int size
= operand_size(insn
, pseudo
);
332 warning(insn
->pos
, "right shift by bigger than source value");
333 return replace_with_pseudo(insn
, value_pseudo(0));
336 return replace_with_pseudo(insn
, pseudo
);
340 static int simplify_mul_div(struct instruction
*insn
, long long value
)
342 unsigned long long sbit
= 1ULL << (insn
->size
- 1);
343 unsigned long long bits
= sbit
| (sbit
- 1);
346 return replace_with_pseudo(insn
, insn
->src1
);
348 switch (insn
->opcode
) {
352 return replace_with_pseudo(insn
, insn
->src2
);
355 if (!(value
& sbit
)) // positive
360 insn
->opcode
= OP_NEG
;
368 static int compare_opcode(int opcode
, int inverse
)
374 case OP_SET_EQ
: return OP_SET_NE
;
375 case OP_SET_NE
: return OP_SET_EQ
;
377 case OP_SET_LT
: return OP_SET_GE
;
378 case OP_SET_LE
: return OP_SET_GT
;
379 case OP_SET_GT
: return OP_SET_LE
;
380 case OP_SET_GE
: return OP_SET_LT
;
382 case OP_SET_A
: return OP_SET_BE
;
383 case OP_SET_AE
: return OP_SET_B
;
384 case OP_SET_B
: return OP_SET_AE
;
385 case OP_SET_BE
: return OP_SET_A
;
392 static int simplify_seteq_setne(struct instruction
*insn
, long long value
)
394 struct instruction
*def
= insn
->src1
->def
;
399 if (value
!= 0 && value
!= 1)
405 inverse
= (insn
->opcode
== OP_SET_NE
) == value
;
406 opcode
= def
->opcode
;
408 case OP_BINCMP
... OP_BINCMP_END
:
410 // setcc.n %t <- %a, %b
411 // setne.m %r <- %t, $0
413 // setcc.n %t <- %a, %b
414 // setcc.m %r <- %a, $b
415 // and similar for setne/eq ... 0/1
418 remove_usage(insn
->src1
, &insn
->src1
);
419 insn
->opcode
= compare_opcode(opcode
, inverse
);
420 use_pseudo(insn
, src1
, &insn
->src1
);
421 use_pseudo(insn
, src2
, &insn
->src2
);
429 static int simplify_constant_rightside(struct instruction
*insn
)
431 long long value
= insn
->src2
->value
;
433 switch (insn
->opcode
) {
436 return replace_with_pseudo(insn
, insn
->src2
);
437 goto case_neutral_zero
;
441 insn
->opcode
= OP_ADD
;
442 insn
->src2
= value_pseudo(-value
);
447 case OP_OR
: case OP_XOR
:
452 return replace_with_pseudo(insn
, insn
->src1
);
455 return simplify_asr(insn
, insn
->src1
, value
);
457 case OP_MODU
: case OP_MODS
:
459 return replace_with_pseudo(insn
, value_pseudo(0));
462 case OP_DIVU
: case OP_DIVS
:
463 case OP_MULU
: case OP_MULS
:
464 return simplify_mul_div(insn
, value
);
468 return replace_with_pseudo(insn
, insn
->src1
);
472 return replace_with_pseudo(insn
, insn
->src2
);
477 return simplify_seteq_setne(insn
, value
);
482 static int simplify_constant_leftside(struct instruction
*insn
)
484 long long value
= insn
->src1
->value
;
486 switch (insn
->opcode
) {
487 case OP_ADD
: case OP_OR
: case OP_XOR
:
489 return replace_with_pseudo(insn
, insn
->src2
);
493 case OP_LSR
: case OP_ASR
:
495 case OP_MULU
: case OP_MULS
:
497 return replace_with_pseudo(insn
, insn
->src1
);
503 static int simplify_constant_binop(struct instruction
*insn
)
505 /* FIXME! Verify signs and sizes!! */
506 long long left
= insn
->src1
->value
;
507 long long right
= insn
->src2
->value
;
508 unsigned long long ul
, ur
;
509 long long res
, mask
, bits
;
511 mask
= 1ULL << (insn
->size
-1);
512 bits
= mask
| (mask
-1);
521 switch (insn
->opcode
) {
542 if (left
== mask
&& right
== -1)
554 if (left
== mask
&& right
== -1)
584 /* Binary comparison */
620 replace_with_pseudo(insn
, value_pseudo(res
));
624 static int simplify_binop_same_args(struct instruction
*insn
, pseudo_t arg
)
626 switch (insn
->opcode
) {
628 case OP_SET_LT
: case OP_SET_GT
:
629 case OP_SET_B
: case OP_SET_A
:
630 if (Wtautological_compare
)
631 warning(insn
->pos
, "self-comparison always evaluates to false");
634 return replace_with_pseudo(insn
, value_pseudo(0));
637 case OP_SET_LE
: case OP_SET_GE
:
638 case OP_SET_BE
: case OP_SET_AE
:
639 if (Wtautological_compare
)
640 warning(insn
->pos
, "self-comparison always evaluates to true");
641 return replace_with_pseudo(insn
, value_pseudo(1));
645 return replace_with_pseudo(insn
, arg
);
649 remove_usage(arg
, &insn
->src2
);
650 insn
->src2
= value_pseudo(0);
651 insn
->opcode
= OP_SET_NE
;
661 static int simplify_binop(struct instruction
*insn
)
663 if (dead_insn(insn
, &insn
->src1
, &insn
->src2
, NULL
))
665 if (constant(insn
->src1
)) {
666 if (constant(insn
->src2
))
667 return simplify_constant_binop(insn
);
668 return simplify_constant_leftside(insn
);
670 if (constant(insn
->src2
))
671 return simplify_constant_rightside(insn
);
672 if (insn
->src1
== insn
->src2
)
673 return simplify_binop_same_args(insn
, insn
->src1
);
677 static void switch_pseudo(struct instruction
*insn1
, pseudo_t
*pp1
, struct instruction
*insn2
, pseudo_t
*pp2
)
679 pseudo_t p1
= *pp1
, p2
= *pp2
;
681 use_pseudo(insn1
, p2
, pp1
);
682 use_pseudo(insn2
, p1
, pp2
);
683 remove_usage(p1
, pp1
);
684 remove_usage(p2
, pp2
);
687 static int canonical_order(pseudo_t p1
, pseudo_t p2
)
689 /* symbol/constants on the right */
690 if (p1
->type
== PSEUDO_VAL
)
691 return p2
->type
== PSEUDO_VAL
;
693 if (p1
->type
== PSEUDO_SYM
)
694 return p2
->type
== PSEUDO_SYM
|| p2
->type
== PSEUDO_VAL
;
699 static int simplify_commutative_binop(struct instruction
*insn
)
701 if (!canonical_order(insn
->src1
, insn
->src2
)) {
702 switch_pseudo(insn
, &insn
->src1
, insn
, &insn
->src2
);
708 static inline int simple_pseudo(pseudo_t pseudo
)
710 return pseudo
->type
== PSEUDO_VAL
|| pseudo
->type
== PSEUDO_SYM
;
713 static int simplify_associative_binop(struct instruction
*insn
)
715 struct instruction
*def
;
716 pseudo_t pseudo
= insn
->src1
;
718 if (!simple_pseudo(insn
->src2
))
720 if (pseudo
->type
!= PSEUDO_REG
)
725 if (def
->opcode
!= insn
->opcode
)
727 if (!simple_pseudo(def
->src2
))
729 if (ptr_list_size((struct ptr_list
*)def
->target
->users
) != 1)
731 switch_pseudo(def
, &def
->src1
, insn
, &insn
->src2
);
735 static int simplify_constant_unop(struct instruction
*insn
)
737 long long val
= insn
->src1
->value
;
740 switch (insn
->opcode
) {
750 mask
= 1ULL << (insn
->size
-1);
751 res
&= mask
| (mask
-1);
753 replace_with_pseudo(insn
, value_pseudo(res
));
757 static int simplify_unop(struct instruction
*insn
)
759 if (dead_insn(insn
, &insn
->src1
, NULL
, NULL
))
761 if (constant(insn
->src1
))
762 return simplify_constant_unop(insn
);
764 switch (insn
->opcode
) {
765 struct instruction
*def
;
768 def
= insn
->src
->def
;
769 if (def
&& def
->opcode
== OP_NOT
)
770 return replace_with_pseudo(insn
, def
->src
);
773 def
= insn
->src
->def
;
774 if (def
&& def
->opcode
== OP_NEG
)
775 return replace_with_pseudo(insn
, def
->src
);
783 static int simplify_one_memop(struct instruction
*insn
, pseudo_t orig
)
785 pseudo_t addr
= insn
->src
;
788 if (addr
->type
== PSEUDO_REG
) {
789 struct instruction
*def
= addr
->def
;
790 if (def
->opcode
== OP_SYMADDR
&& def
->src
) {
791 kill_use(&insn
->src
);
792 use_pseudo(insn
, def
->src
, &insn
->src
);
793 return REPEAT_CSE
| REPEAT_SYMBOL_CLEANUP
;
795 if (def
->opcode
== OP_ADD
) {
815 warning(insn
->pos
, "crazy programmer");
817 insn
->offset
+= off
->value
;
818 use_pseudo(insn
, new, &insn
->src
);
819 remove_usage(addr
, &insn
->src
);
820 return REPEAT_CSE
| REPEAT_SYMBOL_CLEANUP
;
824 * We walk the whole chain of adds/subs backwards. That's not
825 * only more efficient, but it allows us to find loops.
827 static int simplify_memop(struct instruction
*insn
)
830 pseudo_t orig
= insn
->src
;
833 one
= simplify_one_memop(insn
, orig
);
839 static long long get_cast_value(long long val
, int old_size
, int new_size
, int sign
)
843 if (sign
&& new_size
> old_size
) {
844 mask
= 1 << (old_size
-1);
846 val
|= ~(mask
| (mask
-1));
848 mask
= 1 << (new_size
-1);
849 return val
& (mask
| (mask
-1));
852 static int simplify_cast(struct instruction
*insn
)
854 struct symbol
*orig_type
;
858 if (dead_insn(insn
, &insn
->src
, NULL
, NULL
))
861 orig_type
= insn
->orig_type
;
865 /* Keep casts with pointer on either side (not only case of OP_PTRCAST) */
866 if (is_ptr_type(orig_type
) || is_ptr_type(insn
->type
))
869 orig_size
= orig_type
->bit_size
;
873 /* A cast of a constant? */
875 int sign
= orig_type
->ctype
.modifiers
& MOD_SIGNED
;
876 long long val
= get_cast_value(src
->value
, orig_size
, size
, sign
);
877 src
= value_pseudo(val
);
881 /* A cast of a "and" might be a no-op.. */
882 if (src
->type
== PSEUDO_REG
) {
883 struct instruction
*def
= src
->def
;
884 if (def
->opcode
== OP_AND
&& def
->size
>= size
) {
885 pseudo_t val
= def
->src2
;
886 if (val
->type
== PSEUDO_VAL
) {
887 unsigned long long value
= val
->value
;
888 if (!(value
>> (size
-1)))
894 if (size
== orig_size
) {
895 int op
= (orig_type
->ctype
.modifiers
& MOD_SIGNED
) ? OP_SCAST
: OP_CAST
;
896 if (insn
->opcode
== op
)
903 return replace_with_pseudo(insn
, src
);
906 static int simplify_select(struct instruction
*insn
)
908 pseudo_t cond
, src1
, src2
;
910 if (dead_insn(insn
, &insn
->src1
, &insn
->src2
, &insn
->src3
))
916 if (constant(cond
) || src1
== src2
) {
917 pseudo_t
*kill
, take
;
918 kill_use(&insn
->src1
);
919 take
= cond
->value
? src1
: src2
;
920 kill
= cond
->value
? &insn
->src3
: &insn
->src2
;
922 replace_with_pseudo(insn
, take
);
925 if (constant(src1
) && constant(src2
)) {
926 long long val1
= src1
->value
;
927 long long val2
= src2
->value
;
929 /* The pair 0/1 is special - replace with SETNE/SETEQ */
930 if ((val1
| val2
) == 1) {
931 int opcode
= OP_SET_EQ
;
936 insn
->opcode
= opcode
;
937 /* insn->src1 is already cond */
938 insn
->src2
= src1
; /* Zero */
945 static int is_in_range(pseudo_t src
, long long low
, long long high
)
952 return value
>= low
&& value
<= high
;
958 static int simplify_range(struct instruction
*insn
)
960 pseudo_t src1
, src2
, src3
;
965 if (src2
->type
!= PSEUDO_VAL
|| src3
->type
!= PSEUDO_VAL
)
967 if (is_in_range(src1
, src2
->value
, src3
->value
)) {
968 kill_instruction(insn
);
975 * Simplify "set_ne/eq $0 + br"
977 static int simplify_cond_branch(struct instruction
*br
, pseudo_t cond
, struct instruction
*def
, pseudo_t
*pp
)
979 use_pseudo(br
, *pp
, &br
->cond
);
980 remove_usage(cond
, &br
->cond
);
981 if (def
->opcode
== OP_SET_EQ
) {
982 struct basic_block
*true = br
->bb_true
;
983 struct basic_block
*false = br
->bb_false
;
990 static int simplify_branch(struct instruction
*insn
)
992 pseudo_t cond
= insn
->cond
;
997 /* Constant conditional */
998 if (constant(cond
)) {
999 insert_branch(insn
->bb
, insn
, cond
->value
? insn
->bb_true
: insn
->bb_false
);
1004 if (insn
->bb_true
== insn
->bb_false
) {
1005 struct basic_block
*bb
= insn
->bb
;
1006 struct basic_block
*target
= insn
->bb_false
;
1007 remove_bb_from_list(&target
->parents
, bb
, 1);
1008 remove_bb_from_list(&bb
->children
, target
, 1);
1009 insn
->bb_false
= NULL
;
1010 kill_use(&insn
->cond
);
1015 /* Conditional on a SETNE $0 or SETEQ $0 */
1016 if (cond
->type
== PSEUDO_REG
) {
1017 struct instruction
*def
= cond
->def
;
1019 if (def
->opcode
== OP_SET_NE
|| def
->opcode
== OP_SET_EQ
) {
1020 if (constant(def
->src1
) && !def
->src1
->value
)
1021 return simplify_cond_branch(insn
, cond
, def
, &def
->src2
);
1022 if (constant(def
->src2
) && !def
->src2
->value
)
1023 return simplify_cond_branch(insn
, cond
, def
, &def
->src1
);
1025 if (def
->opcode
== OP_SEL
) {
1026 if (constant(def
->src2
) && constant(def
->src3
)) {
1027 long long val1
= def
->src2
->value
;
1028 long long val2
= def
->src3
->value
;
1029 if (!val1
&& !val2
) {
1030 insert_branch(insn
->bb
, insn
, insn
->bb_false
);
1034 insert_branch(insn
->bb
, insn
, insn
->bb_true
);
1038 struct basic_block
*true = insn
->bb_true
;
1039 struct basic_block
*false = insn
->bb_false
;
1040 insn
->bb_false
= true;
1041 insn
->bb_true
= false;
1043 use_pseudo(insn
, def
->src1
, &insn
->cond
);
1044 remove_usage(cond
, &insn
->cond
);
1048 if (def
->opcode
== OP_CAST
|| def
->opcode
== OP_SCAST
) {
1049 int orig_size
= def
->orig_type
? def
->orig_type
->bit_size
: 0;
1050 if (def
->size
> orig_size
) {
1051 use_pseudo(insn
, def
->src
, &insn
->cond
);
1052 remove_usage(cond
, &insn
->cond
);
1060 static int simplify_switch(struct instruction
*insn
)
1062 pseudo_t cond
= insn
->cond
;
1064 struct multijmp
*jmp
;
1066 if (!constant(cond
))
1068 val
= insn
->cond
->value
;
1070 FOR_EACH_PTR(insn
->multijmp_list
, jmp
) {
1072 if (jmp
->begin
> jmp
->end
)
1074 if (val
>= jmp
->begin
&& val
<= jmp
->end
)
1076 } END_FOR_EACH_PTR(jmp
);
1077 warning(insn
->pos
, "Impossible case statement");
1081 insert_branch(insn
->bb
, insn
, jmp
->target
);
1085 int simplify_instruction(struct instruction
*insn
)
1089 switch (insn
->opcode
) {
1090 case OP_ADD
: case OP_MULS
:
1091 case OP_AND
: case OP_OR
: case OP_XOR
:
1092 case OP_AND_BOOL
: case OP_OR_BOOL
:
1093 if (simplify_binop(insn
))
1095 if (simplify_commutative_binop(insn
))
1097 return simplify_associative_binop(insn
);
1100 case OP_SET_EQ
: case OP_SET_NE
:
1101 if (simplify_binop(insn
))
1103 return simplify_commutative_binop(insn
);
1106 case OP_DIVU
: case OP_DIVS
:
1107 case OP_MODU
: case OP_MODS
:
1109 case OP_LSR
: case OP_ASR
:
1110 case OP_SET_LE
: case OP_SET_GE
:
1111 case OP_SET_LT
: case OP_SET_GT
:
1112 case OP_SET_B
: case OP_SET_A
:
1113 case OP_SET_BE
: case OP_SET_AE
:
1114 return simplify_binop(insn
);
1116 case OP_NOT
: case OP_NEG
:
1117 return simplify_unop(insn
);
1118 case OP_LOAD
: case OP_STORE
:
1119 return simplify_memop(insn
);
1121 if (dead_insn(insn
, NULL
, NULL
, NULL
))
1122 return REPEAT_CSE
| REPEAT_SYMBOL_CLEANUP
;
1123 return replace_with_pseudo(insn
, insn
->symbol
);
1128 return simplify_cast(insn
);
1130 if (dead_insn(insn
, NULL
, NULL
, NULL
)) {
1134 return clean_up_phi(insn
);
1136 if (dead_insn(insn
, &insn
->phi_src
, NULL
, NULL
))
1140 return simplify_select(insn
);
1142 return simplify_branch(insn
);
1144 return simplify_switch(insn
);
1146 return simplify_range(insn
);