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 simplify_constant_rightside(struct instruction
*insn
)
370 long long value
= insn
->src2
->value
;
372 switch (insn
->opcode
) {
375 return replace_with_pseudo(insn
, insn
->src2
);
376 goto case_neutral_zero
;
380 insn
->opcode
= OP_ADD
;
381 insn
->src2
= value_pseudo(-value
);
386 case OP_OR
: case OP_XOR
:
391 return replace_with_pseudo(insn
, insn
->src1
);
394 return simplify_asr(insn
, insn
->src1
, value
);
396 case OP_MODU
: case OP_MODS
:
398 return replace_with_pseudo(insn
, value_pseudo(0));
401 case OP_DIVU
: case OP_DIVS
:
402 case OP_MULU
: case OP_MULS
:
403 return simplify_mul_div(insn
, value
);
407 return replace_with_pseudo(insn
, insn
->src1
);
411 return replace_with_pseudo(insn
, insn
->src2
);
417 static int simplify_constant_leftside(struct instruction
*insn
)
419 long long value
= insn
->src1
->value
;
421 switch (insn
->opcode
) {
422 case OP_ADD
: case OP_OR
: case OP_XOR
:
424 return replace_with_pseudo(insn
, insn
->src2
);
428 case OP_LSR
: case OP_ASR
:
430 case OP_MULU
: case OP_MULS
:
432 return replace_with_pseudo(insn
, insn
->src1
);
438 static int simplify_constant_binop(struct instruction
*insn
)
440 /* FIXME! Verify signs and sizes!! */
441 long long left
= insn
->src1
->value
;
442 long long right
= insn
->src2
->value
;
443 unsigned long long ul
, ur
;
444 long long res
, mask
, bits
;
446 mask
= 1ULL << (insn
->size
-1);
447 bits
= mask
| (mask
-1);
456 switch (insn
->opcode
) {
477 if (left
== mask
&& right
== -1)
489 if (left
== mask
&& right
== -1)
519 /* Binary comparison */
555 replace_with_pseudo(insn
, value_pseudo(res
));
559 static int simplify_binop_same_args(struct instruction
*insn
, pseudo_t arg
)
561 switch (insn
->opcode
) {
563 case OP_SET_LT
: case OP_SET_GT
:
564 case OP_SET_B
: case OP_SET_A
:
567 return replace_with_pseudo(insn
, value_pseudo(0));
570 case OP_SET_LE
: case OP_SET_GE
:
571 case OP_SET_BE
: case OP_SET_AE
:
572 return replace_with_pseudo(insn
, value_pseudo(1));
576 return replace_with_pseudo(insn
, arg
);
580 // simplification is correct only if the operands
581 // have already been compared against zero which
592 static int simplify_binop(struct instruction
*insn
)
594 if (dead_insn(insn
, &insn
->src1
, &insn
->src2
, NULL
))
596 if (constant(insn
->src1
)) {
597 if (constant(insn
->src2
))
598 return simplify_constant_binop(insn
);
599 return simplify_constant_leftside(insn
);
601 if (constant(insn
->src2
))
602 return simplify_constant_rightside(insn
);
603 if (insn
->src1
== insn
->src2
)
604 return simplify_binop_same_args(insn
, insn
->src1
);
608 static void switch_pseudo(struct instruction
*insn1
, pseudo_t
*pp1
, struct instruction
*insn2
, pseudo_t
*pp2
)
610 pseudo_t p1
= *pp1
, p2
= *pp2
;
612 use_pseudo(insn1
, p2
, pp1
);
613 use_pseudo(insn2
, p1
, pp2
);
614 remove_usage(p1
, pp1
);
615 remove_usage(p2
, pp2
);
618 static int canonical_order(pseudo_t p1
, pseudo_t p2
)
620 /* symbol/constants on the right */
621 if (p1
->type
== PSEUDO_VAL
)
622 return p2
->type
== PSEUDO_VAL
;
624 if (p1
->type
== PSEUDO_SYM
)
625 return p2
->type
== PSEUDO_SYM
|| p2
->type
== PSEUDO_VAL
;
630 static int simplify_commutative_binop(struct instruction
*insn
)
632 if (!canonical_order(insn
->src1
, insn
->src2
)) {
633 switch_pseudo(insn
, &insn
->src1
, insn
, &insn
->src2
);
639 static inline int simple_pseudo(pseudo_t pseudo
)
641 return pseudo
->type
== PSEUDO_VAL
|| pseudo
->type
== PSEUDO_SYM
;
644 static int simplify_associative_binop(struct instruction
*insn
)
646 struct instruction
*def
;
647 pseudo_t pseudo
= insn
->src1
;
649 if (!simple_pseudo(insn
->src2
))
651 if (pseudo
->type
!= PSEUDO_REG
)
656 if (def
->opcode
!= insn
->opcode
)
658 if (!simple_pseudo(def
->src2
))
660 if (ptr_list_size((struct ptr_list
*)def
->target
->users
) != 1)
662 switch_pseudo(def
, &def
->src1
, insn
, &insn
->src2
);
666 static int simplify_constant_unop(struct instruction
*insn
)
668 long long val
= insn
->src1
->value
;
671 switch (insn
->opcode
) {
681 mask
= 1ULL << (insn
->size
-1);
682 res
&= mask
| (mask
-1);
684 replace_with_pseudo(insn
, value_pseudo(res
));
688 static int simplify_unop(struct instruction
*insn
)
690 if (dead_insn(insn
, &insn
->src1
, NULL
, NULL
))
692 if (constant(insn
->src1
))
693 return simplify_constant_unop(insn
);
695 switch (insn
->opcode
) {
696 struct instruction
*def
;
699 def
= insn
->src
->def
;
700 if (def
&& def
->opcode
== OP_NOT
)
701 return replace_with_pseudo(insn
, def
->src
);
704 def
= insn
->src
->def
;
705 if (def
&& def
->opcode
== OP_NEG
)
706 return replace_with_pseudo(insn
, def
->src
);
714 static int simplify_one_memop(struct instruction
*insn
, pseudo_t orig
)
716 pseudo_t addr
= insn
->src
;
719 if (addr
->type
== PSEUDO_REG
) {
720 struct instruction
*def
= addr
->def
;
721 if (def
->opcode
== OP_SYMADDR
&& def
->src
) {
722 kill_use(&insn
->src
);
723 use_pseudo(insn
, def
->src
, &insn
->src
);
724 return REPEAT_CSE
| REPEAT_SYMBOL_CLEANUP
;
726 if (def
->opcode
== OP_ADD
) {
746 warning(insn
->pos
, "crazy programmer");
748 insn
->offset
+= off
->value
;
749 use_pseudo(insn
, new, &insn
->src
);
750 remove_usage(addr
, &insn
->src
);
751 return REPEAT_CSE
| REPEAT_SYMBOL_CLEANUP
;
755 * We walk the whole chain of adds/subs backwards. That's not
756 * only more efficient, but it allows us to find loops.
758 static int simplify_memop(struct instruction
*insn
)
761 pseudo_t orig
= insn
->src
;
764 one
= simplify_one_memop(insn
, orig
);
770 static long long get_cast_value(long long val
, int old_size
, int new_size
, int sign
)
774 if (sign
&& new_size
> old_size
) {
775 mask
= 1 << (old_size
-1);
777 val
|= ~(mask
| (mask
-1));
779 mask
= 1 << (new_size
-1);
780 return val
& (mask
| (mask
-1));
783 static int simplify_cast(struct instruction
*insn
)
785 struct symbol
*orig_type
;
789 if (dead_insn(insn
, &insn
->src
, NULL
, NULL
))
792 orig_type
= insn
->orig_type
;
796 /* Keep casts with pointer on either side (not only case of OP_PTRCAST) */
797 if (is_ptr_type(orig_type
) || is_ptr_type(insn
->type
))
800 orig_size
= orig_type
->bit_size
;
804 /* A cast of a constant? */
806 int sign
= orig_type
->ctype
.modifiers
& MOD_SIGNED
;
807 long long val
= get_cast_value(src
->value
, orig_size
, size
, sign
);
808 src
= value_pseudo(val
);
812 /* A cast of a "and" might be a no-op.. */
813 if (src
->type
== PSEUDO_REG
) {
814 struct instruction
*def
= src
->def
;
815 if (def
->opcode
== OP_AND
&& def
->size
>= size
) {
816 pseudo_t val
= def
->src2
;
817 if (val
->type
== PSEUDO_VAL
) {
818 unsigned long long value
= val
->value
;
819 if (!(value
>> (size
-1)))
825 if (size
== orig_size
) {
826 int op
= (orig_type
->ctype
.modifiers
& MOD_SIGNED
) ? OP_SCAST
: OP_CAST
;
827 if (insn
->opcode
== op
)
834 return replace_with_pseudo(insn
, src
);
837 static int simplify_select(struct instruction
*insn
)
839 pseudo_t cond
, src1
, src2
;
841 if (dead_insn(insn
, &insn
->src1
, &insn
->src2
, &insn
->src3
))
847 if (constant(cond
) || src1
== src2
) {
848 pseudo_t
*kill
, take
;
849 kill_use(&insn
->src1
);
850 take
= cond
->value
? src1
: src2
;
851 kill
= cond
->value
? &insn
->src3
: &insn
->src2
;
853 replace_with_pseudo(insn
, take
);
856 if (constant(src1
) && constant(src2
)) {
857 long long val1
= src1
->value
;
858 long long val2
= src2
->value
;
860 /* The pair 0/1 is special - replace with SETNE/SETEQ */
861 if ((val1
| val2
) == 1) {
862 int opcode
= OP_SET_EQ
;
867 insn
->opcode
= opcode
;
868 /* insn->src1 is already cond */
869 insn
->src2
= src1
; /* Zero */
876 static int is_in_range(pseudo_t src
, long long low
, long long high
)
883 return value
>= low
&& value
<= high
;
889 static int simplify_range(struct instruction
*insn
)
891 pseudo_t src1
, src2
, src3
;
896 if (src2
->type
!= PSEUDO_VAL
|| src3
->type
!= PSEUDO_VAL
)
898 if (is_in_range(src1
, src2
->value
, src3
->value
)) {
899 kill_instruction(insn
);
906 * Simplify "set_ne/eq $0 + br"
908 static int simplify_cond_branch(struct instruction
*br
, pseudo_t cond
, struct instruction
*def
, pseudo_t
*pp
)
910 use_pseudo(br
, *pp
, &br
->cond
);
911 remove_usage(cond
, &br
->cond
);
912 if (def
->opcode
== OP_SET_EQ
) {
913 struct basic_block
*true = br
->bb_true
;
914 struct basic_block
*false = br
->bb_false
;
921 static int simplify_branch(struct instruction
*insn
)
923 pseudo_t cond
= insn
->cond
;
928 /* Constant conditional */
929 if (constant(cond
)) {
930 insert_branch(insn
->bb
, insn
, cond
->value
? insn
->bb_true
: insn
->bb_false
);
935 if (insn
->bb_true
== insn
->bb_false
) {
936 struct basic_block
*bb
= insn
->bb
;
937 struct basic_block
*target
= insn
->bb_false
;
938 remove_bb_from_list(&target
->parents
, bb
, 1);
939 remove_bb_from_list(&bb
->children
, target
, 1);
940 insn
->bb_false
= NULL
;
941 kill_use(&insn
->cond
);
946 /* Conditional on a SETNE $0 or SETEQ $0 */
947 if (cond
->type
== PSEUDO_REG
) {
948 struct instruction
*def
= cond
->def
;
950 if (def
->opcode
== OP_SET_NE
|| def
->opcode
== OP_SET_EQ
) {
951 if (constant(def
->src1
) && !def
->src1
->value
)
952 return simplify_cond_branch(insn
, cond
, def
, &def
->src2
);
953 if (constant(def
->src2
) && !def
->src2
->value
)
954 return simplify_cond_branch(insn
, cond
, def
, &def
->src1
);
956 if (def
->opcode
== OP_SEL
) {
957 if (constant(def
->src2
) && constant(def
->src3
)) {
958 long long val1
= def
->src2
->value
;
959 long long val2
= def
->src3
->value
;
960 if (!val1
&& !val2
) {
961 insert_branch(insn
->bb
, insn
, insn
->bb_false
);
965 insert_branch(insn
->bb
, insn
, insn
->bb_true
);
969 struct basic_block
*true = insn
->bb_true
;
970 struct basic_block
*false = insn
->bb_false
;
971 insn
->bb_false
= true;
972 insn
->bb_true
= false;
974 use_pseudo(insn
, def
->src1
, &insn
->cond
);
975 remove_usage(cond
, &insn
->cond
);
979 if (def
->opcode
== OP_CAST
|| def
->opcode
== OP_SCAST
) {
980 int orig_size
= def
->orig_type
? def
->orig_type
->bit_size
: 0;
981 if (def
->size
> orig_size
) {
982 use_pseudo(insn
, def
->src
, &insn
->cond
);
983 remove_usage(cond
, &insn
->cond
);
991 static int simplify_switch(struct instruction
*insn
)
993 pseudo_t cond
= insn
->cond
;
995 struct multijmp
*jmp
;
999 val
= insn
->cond
->value
;
1001 FOR_EACH_PTR(insn
->multijmp_list
, jmp
) {
1003 if (jmp
->begin
> jmp
->end
)
1005 if (val
>= jmp
->begin
&& val
<= jmp
->end
)
1007 } END_FOR_EACH_PTR(jmp
);
1008 warning(insn
->pos
, "Impossible case statement");
1012 insert_branch(insn
->bb
, insn
, jmp
->target
);
1016 int simplify_instruction(struct instruction
*insn
)
1020 switch (insn
->opcode
) {
1021 case OP_ADD
: case OP_MULS
:
1022 case OP_AND
: case OP_OR
: case OP_XOR
:
1023 case OP_AND_BOOL
: case OP_OR_BOOL
:
1024 if (simplify_binop(insn
))
1026 if (simplify_commutative_binop(insn
))
1028 return simplify_associative_binop(insn
);
1031 case OP_SET_EQ
: case OP_SET_NE
:
1032 if (simplify_binop(insn
))
1034 return simplify_commutative_binop(insn
);
1037 case OP_DIVU
: case OP_DIVS
:
1038 case OP_MODU
: case OP_MODS
:
1040 case OP_LSR
: case OP_ASR
:
1041 case OP_SET_LE
: case OP_SET_GE
:
1042 case OP_SET_LT
: case OP_SET_GT
:
1043 case OP_SET_B
: case OP_SET_A
:
1044 case OP_SET_BE
: case OP_SET_AE
:
1045 return simplify_binop(insn
);
1047 case OP_NOT
: case OP_NEG
:
1048 return simplify_unop(insn
);
1049 case OP_LOAD
: case OP_STORE
:
1050 return simplify_memop(insn
);
1052 if (dead_insn(insn
, NULL
, NULL
, NULL
))
1053 return REPEAT_CSE
| REPEAT_SYMBOL_CLEANUP
;
1054 return replace_with_pseudo(insn
, insn
->symbol
);
1059 return simplify_cast(insn
);
1061 if (dead_insn(insn
, NULL
, NULL
, NULL
)) {
1065 return clean_up_phi(insn
);
1067 if (dead_insn(insn
, &insn
->phi_src
, NULL
, NULL
))
1071 return simplify_select(insn
);
1073 return simplify_branch(insn
);
1075 return simplify_switch(insn
);
1077 return simplify_range(insn
);