1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
|
#include "ir.h"
static int
mulk(struct instr *ins, struct block *blk, int *curi)
{
vlong iv = intconval(ins->r);
enum irclass cls = ins->cls;
assert(iv > 1 && "trivial mul not handled by irbinop() ?");
/* This can be generalized to any sequence of shifts and
* adds/subtracts, but whether that's worth it depends on the number of them
* and the microarchitecture.. clang seems to stop after two shifts. Should
* we compute approximate cost of instrs to determine? For now just handle
* the po2 (+/- 1) case */
if (ispo2(iv)) {
/* x * 2^y ==> x << y */
ins->op = Oshl;
ins->r = mkref(RICON, ilog2(iv));
return 1;
} else if (ispo2(iv-1)) {
/* x * 5 ==> (x << 2) + x */
ins->op = Oadd;
ins->r = ins->l;
ins->l = insertinstr(blk, (*curi)++, mkinstr(Oshl, cls, ins->l, mkref(RICON, ilog2(iv-1))));
return 1;
} else if (ispo2(iv+1)) {
/* x * 7 ==> (x << 3) - x */
ins->op = Osub;
ins->r = ins->l;
ins->l = insertinstr(blk, (*curi)++, mkinstr(Oshl, cls, ins->l, mkref(RICON, ilog2(iv+1))));
return 1;
}
return 0;
}
static int
ins(struct instr *ins, struct block *blk, int *curi)
{
int narg = opnarg[ins->op];
if (oisarith(ins->op)) {
union ref r = narg == 1 ? irunop(NULL, ins->op, ins->cls, ins->l)
: irbinop(NULL, ins->op, ins->cls, ins->l, ins->r);
if (r.bits) {
ins->op = Ocopy;
ins->cls = insrescls(*ins);
ins->l = r;
ins->r = NOREF;
return 1;
}
}
enum irclass k = ins->cls;
switch (ins->op) {
case Omul:
if (kisflt(k)) break;
if (isnumcon(ins->l)) rswap(ins->l, ins->r); /* put const in rhs */
if (isintcon(ins->r)) return mulk(ins, blk, curi);
}
return 0;
}
static void
jmpfind(struct block **final, struct block **pblk)
{
struct block **p2 = &final[(*pblk)->id];
if (*p2 && !(*p2)->phi.n) {
jmpfind(final, p2);
*pblk = *p2;
}
}
static void
fillpredsrec(struct block *blk)
{
while (blk && !wasvisited(blk)) {
markvisited(blk);
if (!blk->s1) return;
if (!blk->s1->phi.n) addpred(blk->s1, blk);
if (!blk->s2) {
blk = blk->s1;
continue;
}
if (!blk->s2->phi.n) addpred(blk->s2, blk);
fillpredsrec(blk->s1);
blk = blk->s2;
}
}
static void
fillpreds(struct function *fn)
{
struct block *blk = fn->entry, *next;
do {
if (blk->phi.n) continue;
if (blk->npred > 1)
xbfree(blk->_pred);
blk->npred = 0;
blk->_pred = NULL;
} while ((blk = blk->lnext) != fn->entry);
startbbvisit();
fillpredsrec(fn->entry);
blk = fn->entry->lnext;
do { /* remove dead blocks */
next = blk->lnext;
if (!blk->npred) freeblk(fn, blk);
} while ((blk = next) != fn->entry);
}
void
simpl(struct function *fn)
{
FREQUIRE(FNUSE);
int inschange = 0,
blkchange = 0;
struct block **jmpfinal = allocz(fn->passarena, fn->nblk * sizeof *jmpfinal, 0);
struct block *blk = fn->entry;
fn->curblk = NULL;
do {
for (int i = 0; i < blk->phi.n; ++i) {
}
for (int i = 0; i < blk->ins.n; ++i) {
inschange += ins(&instrtab[blk->ins.p[i]], blk, &i);
}
/* merge blocks:
* @blk:
* ... (1)
* b @s
* @s:
* ... (2)
* b {...}
*=>
* @blk:
* ... (1)
* ... (2)
* b {...}
* */
if (blk != fn->entry) while (blk->s1 && !blk->s2 && blk->s1->npred == 1 && !blk->phi.n) {
struct block *s = blk->s1;
vpushn(&blk->ins, s->ins.p, s->ins.n);
blk->jmp = blk->s1->jmp;
for (int i = 0; i < 2; ++i) {
struct block *ss = (&s->s1)[i];
if (ss) for (int j = 0; j < ss->npred; ++j) {
if (blkpred(ss, j) == s) {
blkpred(ss, j) = blk;
break;
}
}
}
/* breaks use for temps used in s */
if (s->ins.n || s->jmp.arg[0].t == RTMP || s->jmp.arg[1].t == RTMP)
fn->prop &= ~FNUSE;
blk->s1 = s->s1;
blk->s2 = s->s2;
freeblk(fn, s);
}
/* thread jumps.. */
if (!blk->phi.n && !blk->ins.n) {
if (blk->jmp.t == Jb && !blk->s2) {
jmpfind(jmpfinal, &blk->s1);
if (blk->s1 != blk) {
++blkchange;
jmpfinal[blk->id] = blk->s1;
}
}
}
} while ((blk = blk->lnext) != fn->entry);
if (inschange) {
if (!(fn->prop & FNUSE)) filluses(fn);
copyopt(fn);
}
if (blkchange) {
do {
if (blk->s1) jmpfind(jmpfinal, &blk->s1);
if (blk->s2) jmpfind(jmpfinal, &blk->s2);
if (blk->s1 && blk->s1 == blk->s2) {
memset(blk->jmp.arg, 0, sizeof blk->jmp.arg);
blk->s2 = NULL;
}
} while ((blk = blk->lnext) != fn->entry);
fillpreds(fn);
}
}
/* vim:set ts=3 sw=3 expandtab: */
|
