void
ivars(void)
{
char *temp=NULL;
env_overr(&Spice_Exec_Dir, "SPICE_EXEC_DIR");
env_overr(&Spice_Lib_Dir, "SPICE_LIB_DIR");
mkvar(&News_File, Spice_Lib_Dir, "news", "SPICE_NEWS");
mkvar(&Default_MFB_Cap, Spice_Lib_Dir, "mfbcap", "SPICE_MFBCAP");
mkvar(&Help_Path, Spice_Lib_Dir, "helpdir", "SPICE_HELP_DIR");
mkvar(&Lib_Path, Spice_Lib_Dir, "scripts", "SPICE_SCRIPTS");
mkvar(&Spice_Path, Spice_Exec_Dir, "ngspice", "SPICE_PATH");
env_overr(&Spice_Host, "SPICE_HOST");
env_overr(&Bug_Addr, "SPICE_BUGADDR");
env_overr(&Def_Editor, "SPICE_EDITOR");
env_overr(&temp, "SPICE_ASCIIRAWFILE");
if(temp)
AsciiRawFile = atoi(temp);
}
extern void varpop(Push *push) {
Var *var;
assert(pushlist == push);
assert(rootlist == &push->defnroot);
assert(rootlist->next == &push->nameroot);
if (isexported(push->name))
isdirty = TRUE;
push->defn = callsettor(push->name, push->defn);
var = dictget(vars, push->name);
if (var != NULL)
if (push->defn != NULL) {
var->defn = push->defn;
var->flags = push->flags;
var->env = NULL;
} else
vars = dictput(vars, push->name, NULL);
else if (push->defn != NULL) {
var = mkvar(NULL);
var->defn = push->defn;
var->flags = push->flags;
vars = dictput(vars, push->name, var);
}
pushlist = pushlist->next;
rootlist = rootlist->next->next;
}
extern void vardef(char *name, Binding *binding, List *defn) {
Var *var;
validatevar(name);
for (; binding != NULL; binding = binding->next)
if (streq(name, binding->name)) {
binding->defn = defn;
rebound = TRUE;
return;
}
RefAdd(name);
defn = callsettor(name, defn);
if (isexported(name))
isdirty = TRUE;
var = dictget(vars, name);
if (var != NULL)
if (defn != NULL) {
var->defn = defn;
var->env = NULL;
var->flags = hasbindings(defn) ? var_hasbindings : 0;
} else
vars = dictput(vars, name, NULL);
else if (defn != NULL) {
var = mkvar(defn);
vars = dictput(vars, name, var);
}
RefRemove(name);
}
extern void varpush(Push *push, char *name, List *defn) {
Var *var;
validatevar(name);
push->name = name;
push->nameroot.next = rootlist;
push->nameroot.p = (void **) &push->name;
rootlist = &push->nameroot;
if (isexported(name))
isdirty = TRUE;
defn = callsettor(name, defn);
var = dictget(vars, push->name);
if (var == NULL) {
push->defn = NULL;
push->flags = 0;
var = mkvar(defn);
vars = dictput(vars, push->name, var);
} else {
push->defn = var->defn;
push->flags = var->flags;
var->defn = defn;
var->env = NULL;
var->flags = hasbindings(defn) ? var_hasbindings : 0;
}
push->next = pushlist;
pushlist = push;
push->defnroot.next = rootlist;
push->defnroot.p = (void **) &push->defn;
rootlist = &push->defnroot;
}
/* determine whether we should determine the next value of a variable.
also decide which variables to registerize. */
static void
trackneed(void)
{
SemVar *v;
int ch, i, j, o;
for(j = 0; j < nvars; j++) {
v = vars[j];
if(v->idx == 0 && v->prime) {
if(v->sym->semc[0]->def != 0 && v->sym->semc[1]->def != 0) {
error(v->sym, "'%s' both primed and unprimed defined", v->sym->name);
continue;
}
if((v->sym->opt & OPTREG) != 0 && v->sym->semc[0]->def != 0) {
if((v->sym->semc[0]->flags & SVCANNX) == 0) {
error(v->sym, "'%s' cannot be register", v->sym->name);
v->sym->opt &= ~OPTREG;
} else
v->sym->semc[0]->flags |= SVNEEDNX | SVDELDEF | SVREG;
}
if((v->sym->opt & (OPTOUT|OPTWIRE)) == OPTOUT && v->sym->semc[0]->def != 0 && (v->sym->semc[0]->flags & SVCANNX) != 0)
v->sym->semc[0]->flags |= SVNEEDNX | SVDELDEF | SVREG;
if(v->sym->semc[1]->def != 0) {
if((v->sym->opt & OPTWIRE) != 0) {
error(v->sym, "'%s' cannot be wire", v->sym->name);
v->sym->opt &= ~OPTWIRE;
} else
v->sym->semc[0]->flags |= SVREG;
}
}
}
/* propagate SVNEEDNX to all variables we are dependent on */
do {
ch = 0;
for(j = 0; j < nvars; j++) {
v = vars[j];
if((v->flags & SVNEEDNX) == 0) continue;
if(v->deps == nil) continue;
for(i = 0; i < v->deps->n; i++) {
o = v->deps->p[i]->flags & SVNEEDNX;
v->deps->p[i]->flags |= SVNEEDNX;
ch += o == 0;
}
}
} while(ch > 0);
for(j = 0; j < nvars; j++) {
v = vars[j];
if((v->flags & SVREG) != 0)
v->tnext = v->sym->semc[1];
else if((v->flags & SVNEEDNX) != 0)
v->tnext = mkvar(v->sym, 1);
}
}
/* add initialization statements to blocks as needed */
static void
sinitbuild(SemBlock *b)
{
SemBlock *yes, *no, *then;
SemInit *si;
ASTNode *p;
SemVar *nv, *v;
SemTrigger *t;
for(t = sinits; t != nil; t = t->next) {
yes = newblock();
no = newblock();
then = newblock();
yes->cont = node(ASTBLOCK, nil);
then->phi = node(ASTBLOCK, nil);
b->jump = node(ASTIF, t->trigger, node(ASTSEMGOTO, yes), node(ASTSEMGOTO, no));
yes->jump = node(ASTSEMGOTO, then);
no->jump = node(ASTSEMGOTO, then);
mkftlist(b, 1, yes, no, nil);
mkftlist(yes, 0, b, nil);
mkftlist(yes, 1, then, nil);
mkftlist(no, 0, b, nil);
mkftlist(no, 1, then, nil);
mkftlist(then, 0, yes, no, nil);
for(si = t->first; si != nil; si = si->tnext) {
v = mkvar(si->var->sym, 1);
if(--si->var->nsinits > 0)
nv = mkvar(si->var->sym, 1);
else
nv = si->var->sym->semc[1];
yes->cont->nl = nlcat(yes->cont->nl, nl(node(ASTASS, OPNOP, node(ASTSSA, v), si->val)));
p = node(ASTPHI);
p->nl = nls(node(ASTSSA, v), node(ASTSSA, ssaget(sinitvars, v->sym, 1)), nil);
then->phi->nl = nlcat(then->phi->nl, nl(node(ASTASS, OPNOP, node(ASTSSA, nv), p)));
defsadd(sinitvars, nv, 1);
}
b = then;
}
}
/* find all variables in a statement that we would like to initialize */
static Nodes *
sinitblock(ASTNode *n)
{
SemVar *v, *nv;
SemInit *s;
if(n->t != ASTASS || n->n1->t != ASTSSA) return nl(n);
v = n->n1->semv;
if(v != v->sym->semc[1]) return nl(n);
for(s = v->sym->semc[0]->init; s != nil; s = s->vnext)
if(s->type == SISYNC)
break;
if(s == nil) return nl(n);
nv = mkvar(v->sym, v->prime);
n->n1 = node(ASTSSA, nv);
defsadd(sinitvars, nv, 1);
return nl(n);
}
static void
setoutvar(void)
{
Type *t;
Node *n;
Addr a;
Iter save;
Bits bit;
int z;
t = structfirst(&save, getoutarg(curfn->type));
while(t != T) {
n = nodarg(t, 1);
a = zprog.from;
naddr(n, &a, 0);
bit = mkvar(R, &a);
for(z=0; z<BITS; z++)
ovar.b[z] |= bit.b[z];
t = structnext(&save);
}
//if(bany(&ovar))
//print("ovars = %Q\n", ovar);
}
Bits
mkvar(Reg *r, Addr *a)
{
Var *v;
int i, t, n, et, z;
int32 o;
Bits bit;
LSym *s;
/*
* mark registers used
*/
t = a->type;
r->regu |= doregbits(t);
r->regu |= doregbits(a->index);
switch(t) {
default:
goto none;
case D_ADDR:
a->type = a->index;
bit = mkvar(r, a);
for(z=0; z<BITS; z++)
addrs.b[z] |= bit.b[z];
a->type = t;
goto none;
case D_EXTERN:
case D_STATIC:
case D_PARAM:
case D_AUTO:
n = t;
break;
}
s = a->sym;
if(s == nil)
goto none;
if(s->name[0] == '.')
goto none;
et = a->etype;
o = a->offset;
v = var;
for(i=0; i<nvar; i++) {
if(s == v->sym)
if(n == v->name)
if(o == v->offset)
goto out;
v++;
}
if(nvar >= NVAR)
fatal(Z, "variable not optimized: %s", s->name);
i = nvar;
nvar++;
v = &var[i];
v->sym = s;
v->offset = o;
v->name = n;
v->etype = et;
if(debug['R'])
print("bit=%2d et=%2d %D\n", i, et, a);
out:
bit = blsh(i);
if(n == D_EXTERN || n == D_STATIC)
for(z=0; z<BITS; z++)
externs.b[z] |= bit.b[z];
if(n == D_PARAM)
for(z=0; z<BITS; z++)
params.b[z] |= bit.b[z];
if(v->etype != et || !(typechlpfd[et] || typev[et])) /* funny punning */
for(z=0; z<BITS; z++)
addrs.b[z] |= bit.b[z];
return bit;
none:
return zbits;
}
/* scan: parse the contents of fp, expecting the input side (phase) of a rule
if ph=='i', or the output side if ph=='o' */
List scan (FILE *fp, int ph) {
int c = 0; /* Current char on fp */
int nb = 0, nw = 0; /* Position in rule (bit/word) */
Symbol csym = 0L; /* Current variable being scanned */
T f_msk=(T)0, f_val=(T)0; /* Current constant mask and value */
List pats = 0L; /* All the patterns (words) for this phase */
List vars = 0L; /* All the vars for a given pattern */
while ((c=getc(fp)) != EOF) {
if (debugp)
fprintf(stdout, "/* bit: %c msk: 0x%x val: 0x%x */\n",
c, (unsigned)f_msk, (unsigned)f_val);
/* Inputs end with '=', outputs with '+' */
if ((c == '=' && ph == 'i') || (c == '+' && ph == 'o'))
break;
switch(c) {
case ' ': case '\f': case '\n': case '\r': case '\t': case '\v':
continue;
case '#': /* Comment character */
while ((c=getc(fp)) != EOF && c != '\n');
if (c == EOF && (nb > 0 || (ph == 'i' && nw > 0)))
error(stringf("Unexpected end of file reading patterns at "
"rule %d, word %c%d, bit %d", nr,ph,nw,nb));
continue;
case '-': /* Denotes body of a variable */
if (!csym) /* We must be defining a variable */
error(stringf("'-' appears with no leading symbol "
"at rule %d, word %c%d, bit %d", nr, ph, nw, nb));
else { /* Extend the variable's right boundary */
++csym->r; assert(csym->r == nb);
}
break;
case '0': case '1': /* Constant characters */
if (csym) { /* End any variable definition */
vars = mkvar(csym, vars); csym = 0L;
}
f_msk |= 1; f_val |= (c-'0');
break;
default:
c = tolower(c); /* Variables named by a-z, case insensitive */
if (c >= 'a' && c <= 'z') {
char *s = stringf("%c", c);
if (csym)
vars = mkvar(csym, vars);
if (!lookup(s, stab) && ph == 'o')
error(stringf("Symbol '%c' used with no prior "
"definition at rule %d, word o%d",
c, nr, nw));
csym = install(s, stab);
csym->l = csym->r = nb;
} else
error(stringf("Illegal character in rule file at rule %d, "
"word %c%d, bit %d", nr, ph, nw, nb));
}
if (nb == SZ-1) { /* End of a pattern (word): append this info */
pat_t *pat; /* to list of patterns, and reset current */
if (csym) { /* pattern state (masks, vars, bit count) */
vars = mkvar(csym, vars); csym = 0L;
}
NEW(pat, 1, ARENA0);
pat->f_msk = f_msk; f_msk = (T)0;
pat->f_val = f_val; f_val = (T)0;
pat->nv = l_length(vars);
if (debugp)
fprintf(stdout, "/* msk: 0x%x, val: 0x%x, nv: 0x%x */\n",
(unsigned)pat->f_msk, (unsigned)pat->f_val,
pat->nv);
l_ltov(pat->v, var_t *, vars, ARENA0);
pats = l_append(pat, pats, ARENA0);
++nw; nb = 0; vars = 0L;
} else { /* Still more to go: move on to next bit */
++nb; f_msk <<= 1; f_val <<= 1;
void
regopt(Prog *firstp)
{
Reg *r, *r1;
Prog *p;
int i, z, nr;
uint32 vreg;
Bits bit;
if(first == 0) {
fmtinstall('Q', Qconv);
}
fixjmp(firstp);
first++;
if(debug['K']) {
if(first != 13)
return;
// debug['R'] = 2;
// debug['P'] = 2;
print("optimizing %S\n", curfn->nname->sym);
}
// count instructions
nr = 0;
for(p=firstp; p!=P; p=p->link)
nr++;
// if too big dont bother
if(nr >= 10000) {
// print("********** %S is too big (%d)\n", curfn->nname->sym, nr);
return;
}
r1 = R;
firstr = R;
lastr = R;
/*
* control flow is more complicated in generated go code
* than in generated c code. define pseudo-variables for
* registers, so we have complete register usage information.
*/
nvar = NREGVAR;
memset(var, 0, NREGVAR*sizeof var[0]);
for(i=0; i<NREGVAR; i++)
var[i].node = newname(lookup(regname[i]));
regbits = RtoB(REGSP)|RtoB(REGLINK)|RtoB(REGPC);
for(z=0; z<BITS; z++) {
externs.b[z] = 0;
params.b[z] = 0;
consts.b[z] = 0;
addrs.b[z] = 0;
ovar.b[z] = 0;
}
// build list of return variables
setoutvar();
/*
* pass 1
* build aux data structure
* allocate pcs
* find use and set of variables
*/
nr = 0;
for(p=firstp; p != P; p = p->link) {
switch(p->as) {
case ADATA:
case AGLOBL:
case ANAME:
case ASIGNAME:
continue;
}
r = rega();
nr++;
if(firstr == R) {
firstr = r;
lastr = r;
} else {
lastr->link = r;
r->p1 = lastr;
lastr->s1 = r;
lastr = r;
}
r->prog = p;
p->regp = r;
r1 = r->p1;
if(r1 != R) {
switch(r1->prog->as) {
case ARET:
case AB:
case ARFE:
r->p1 = R;
r1->s1 = R;
}
}
/*
* left side always read
*/
bit = mkvar(r, &p->from);
for(z=0; z<BITS; z++)
r->use1.b[z] |= bit.b[z];
/*
* middle always read when present
*/
if(p->reg != NREG) {
if(p->from.type != D_FREG)
r->use1.b[0] |= RtoB(p->reg);
else
r->use1.b[0] |= FtoB(p->reg);
}
/*
* right side depends on opcode
*/
bit = mkvar(r, &p->to);
if(bany(&bit))
switch(p->as) {
default:
yyerror("reg: unknown op: %A", p->as);
break;
/*
* right side read
*/
case ATST:
case ATEQ:
case ACMP:
case ACMN:
case ACMPD:
case ACMPF:
rightread:
for(z=0; z<BITS; z++)
r->use2.b[z] |= bit.b[z];
break;
/*
* right side read or read+write, depending on middle
* ADD x, z => z += x
* ADD x, y, z => z = x + y
*/
case AADD:
case AAND:
case AEOR:
case ASUB:
case ARSB:
case AADC:
case ASBC:
case ARSC:
case AORR:
case ABIC:
case ASLL:
case ASRL:
case ASRA:
case AMUL:
case AMULU:
case ADIV:
case AMOD:
case AMODU:
case ADIVU:
if(p->reg != NREG)
goto rightread;
// fall through
/*
* right side read+write
*/
case AADDF:
case AADDD:
case ASUBF:
case ASUBD:
case AMULF:
case AMULD:
case ADIVF:
case ADIVD:
case AMULA:
case AMULAL:
case AMULALU:
for(z=0; z<BITS; z++) {
r->use2.b[z] |= bit.b[z];
r->set.b[z] |= bit.b[z];
}
break;
/*
* right side write
*/
case ANOP:
case AMOVB:
case AMOVBU:
case AMOVD:
case AMOVDF:
case AMOVDW:
case AMOVF:
case AMOVFW:
case AMOVH:
case AMOVHU:
case AMOVW:
case AMOVWD:
case AMOVWF:
case AMVN:
case AMULL:
case AMULLU:
if((p->scond & C_SCOND) != C_SCOND_NONE)
for(z=0; z<BITS; z++)
r->use2.b[z] |= bit.b[z];
for(z=0; z<BITS; z++)
r->set.b[z] |= bit.b[z];
break;
/*
* funny
*/
case ABL:
setaddrs(bit);
break;
}
if(p->as == AMOVM) {
z = p->to.offset;
if(p->from.type == D_CONST)
z = p->from.offset;
for(i=0; z; i++) {
if(z&1)
regbits |= RtoB(i);
z >>= 1;
}
}
}
void
regopt(Prog *firstp)
{
Reg *r, *r1;
Prog *p;
int i, z, nr;
uint32 vreg;
Bits bit;
if(first == 0) {
fmtinstall('Q', Qconv);
}
first++;
if(debug['K']) {
if(first != 13)
return;
// debug['R'] = 2;
// debug['P'] = 2;
print("optimizing %S\n", curfn->nname->sym);
}
// count instructions
nr = 0;
for(p=firstp; p!=P; p=p->link)
nr++;
// if too big dont bother
if(nr >= 10000) {
// print("********** %S is too big (%d)\n", curfn->nname->sym, nr);
return;
}
r1 = R;
firstr = R;
lastr = R;
nvar = 0;
regbits = RtoB(REGSP)|RtoB(REGLINK)|RtoB(REGPC);
for(z=0; z<BITS; z++) {
externs.b[z] = 0;
params.b[z] = 0;
consts.b[z] = 0;
addrs.b[z] = 0;
ovar.b[z] = 0;
}
// build list of return variables
setoutvar();
/*
* pass 1
* build aux data structure
* allocate pcs
* find use and set of variables
*/
nr = 0;
for(p=firstp; p != P; p = p->link) {
switch(p->as) {
case ADATA:
case AGLOBL:
case ANAME:
case ASIGNAME:
continue;
}
r = rega();
nr++;
if(firstr == R) {
firstr = r;
lastr = r;
} else {
lastr->link = r;
r->p1 = lastr;
lastr->s1 = r;
lastr = r;
}
r->prog = p;
p->regp = r;
r1 = r->p1;
if(r1 != R) {
switch(r1->prog->as) {
case ARET:
case AB:
case ARFE:
r->p1 = R;
r1->s1 = R;
}
}
/*
* left side always read
*/
bit = mkvar(r, &p->from);
for(z=0; z<BITS; z++)
r->use1.b[z] |= bit.b[z];
/*
* right side depends on opcode
*/
bit = mkvar(r, &p->to);
if(bany(&bit))
switch(p->as) {
default:
yyerror("reg: unknown op: %A", p->as);
break;
/*
* right side write
*/
case ANOP:
case AMOVB:
case AMOVBU:
case AMOVH:
case AMOVHU:
case AMOVW:
case AMOVF:
case AMOVD:
for(z=0; z<BITS; z++)
r->set.b[z] |= bit.b[z];
break;
/*
* funny
*/
case ABL:
setaddrs(bit);
break;
}
if(p->as == AMOVM) {
z = p->to.offset;
if(p->from.type == D_CONST)
z = p->from.offset;
for(i=0; z; i++) {
if(z&1)
regbits |= RtoB(i);
z >>= 1;
}
}
}
Bits
mkvar(Reg *r, Adr *a, int isro)
{
Var *v;
int i, t, n, et, z;
int32_t o;
Bits bit;
Sym *s;
/*
* mark registers used
*/
t = a->type;
r->regu |= doregbits(t);
r->regu |= doregbits(a->index);
et = a->etype;
switch(t) {
default:
goto none;
case D_INDIR+D_GS:
if(!isro || 1)
goto none;
n = t;
{static Sym er; a->sym = &er;}
a->sym->name = "$extreg";
break;
case D_ADDR:
a->type = a->index;
bit = mkvar(r, a, 0);
for(z=0; z<BITS; z++)
addrs.b[z] |= bit.b[z];
a->type = t;
goto none;
case D_EXTERN:
case D_STATIC:
case D_PARAM:
case D_AUTO:
n = t;
break;
}
s = a->sym;
if(s == S)
goto none;
if(s->name[0] == '.')
goto none;
o = a->offset;
v = var;
for(i=0; i<nvar; i++) {
if(s == v->sym)
if(n == v->name)
if(o == v->offset)
goto out;
v++;
}
if(nvar >= NVAR) {
if(debug['w'] > 1 && s)
warn(Z, "variable not optimized: %s", s->name);
goto none;
}
i = nvar;
nvar++;
v = &var[i];
v->sym = s;
v->offset = o;
v->name = n;
v->etype = et;
if(debug['R'])
print("bit=%2d et=%2d %D\n", i, et, a);
out:
bit = blsh(i);
if(n == D_EXTERN || n == D_STATIC)
for(z=0; z<BITS; z++)
externs.b[z] |= bit.b[z];
if(n == D_PARAM)
for(z=0; z<BITS; z++)
params.b[z] |= bit.b[z];
if(v->etype != et || !typechlpfd[et]) /* funny punning */
for(z=0; z<BITS; z++)
addrs.b[z] |= bit.b[z];
return bit;
none:
return zbits;
}
Bits
mkvar(Reg *r, Adr *a)
{
Var *v;
int i, t, n, et, z, w, flag;
int32 o;
Bits bit;
Sym *s;
/*
* mark registers used
*/
t = a->type;
if(r != R) {
r->regu |= doregbits(t);
r->regu |= doregbits(a->index);
}
switch(t) {
default:
goto none;
case D_ADDR:
a->type = a->index;
bit = mkvar(r, a);
for(z=0; z<BITS; z++)
addrs.b[z] |= bit.b[z];
a->type = t;
ostats.naddr++;
goto none;
case D_EXTERN:
case D_STATIC:
case D_PARAM:
case D_AUTO:
n = t;
break;
}
s = a->sym;
if(s == S)
goto none;
if(s->name[0] == '.')
goto none;
et = a->etype;
o = a->offset;
w = a->width;
v = var;
flag = 0;
for(i=0; i<nvar; i++) {
if(s == v->sym)
if(n == v->name) {
// if it is the same, use it
if(v->etype == et)
if(v->width == w)
if(v->offset == o)
goto out;
// if it overlaps, set max
// width and dont registerize
if(overlap(v, o, w))
flag = 1;
}
v++;
}
if(flag)
goto none;
switch(et) {
case 0:
case TFUNC:
case TARRAY:
goto none;
}
if(nvar >= NVAR) {
if(debug['w'] > 1 && s)
fatal("variable not optimized: %D", a);
goto none;
}
i = nvar;
nvar++;
v = &var[i];
v->sym = s;
v->offset = o;
v->name = n;
v->gotype = a->gotype;
v->etype = et;
v->width = w;
if(debug['R'])
print("bit=%2d et=%2d w=%d %D\n", i, et, w, a);
ostats.nvar++;
out:
bit = blsh(i);
// funny punning
if(v->etype != et) {
if(debug['R'])
print("pun et=%d/%d w=%d/%d o=%d/%d %D\n",
v->etype, et,
v->width, w,
v->offset, o, a);
for(z=0; z<BITS; z++)
addrs.b[z] |= bit.b[z];
goto none;
}
if(n == D_EXTERN || n == D_STATIC)
for(z=0; z<BITS; z++)
externs.b[z] |= bit.b[z];
if(n == D_PARAM)
for(z=0; z<BITS; z++)
params.b[z] |= bit.b[z];
return bit;
none:
return zbits;
}
void
readtext(char *s)
{
Dir *d;
Lsym *l;
Value *v;
Symbol sym;
ulong length;
extern Machdata mipsmach;
if(mtype != 0){
symmap = newmap(0, 1);
if(symmap == 0)
print("%s: (error) loadmap: cannot make symbol map\n", argv0);
length = 1<<24;
d = dirfstat(text);
if(d != nil) {
length = d->length;
free(d);
}
setmap(symmap, text, 0, length, 0, "binary");
free(d);
return;
}
machdata = &mipsmach;
if(!crackhdr(text, &fhdr)) {
print("can't decode file header\n");
return;
}
symmap = loadmap(0, text, &fhdr);
if(symmap == 0)
print("%s: (error) loadmap: cannot make symbol map\n", argv0);
if(syminit(text, &fhdr) < 0) {
print("%s: (error) syminit: %r\n", argv0);
return;
}
print("%s:%s\n\n", s, fhdr.name);
if(mach->sbreg && lookup(0, mach->sbreg, &sym)) {
mach->sb = sym.value;
l = enter("SB", Tid);
l->v->vstore.fmt = 'X';
l->v->vstore.u0.sival = mach->sb;
l->v->type = TINT;
l->v->set = 1;
}
l = mkvar("objtype");
v = l->v;
v->vstore.fmt = 's';
v->set = 1;
v->vstore.u0.sstring = strnode(mach->name);
v->type = TSTRING;
l = mkvar("textfile");
v = l->v;
v->vstore.fmt = 's';
v->set = 1;
v->vstore.u0.sstring = strnode(s);
v->type = TSTRING;
machbytype(fhdr.type);
varreg();
}
void
regopt(Prog *firstp)
{
Reg *r, *r1;
Prog *p;
Graph *g;
int i, z;
uint32 vreg;
Bits bit;
ProgInfo info;
if(first) {
fmtinstall('Q', Qconv);
first = 0;
}
fixjmp(firstp);
mergetemp(firstp);
/*
* control flow is more complicated in generated go code
* than in generated c code. define pseudo-variables for
* registers, so we have complete register usage information.
*/
nvar = NREGVAR;
memset(var, 0, NREGVAR*sizeof var[0]);
for(i=0; i<NREGVAR; i++) {
if(regnodes[i] == N)
regnodes[i] = newname(lookup(regname[i]));
var[i].node = regnodes[i];
}
regbits = RtoB(REGSP)|RtoB(REGLINK)|RtoB(REGPC);
for(z=0; z<BITS; z++) {
externs.b[z] = 0;
params.b[z] = 0;
consts.b[z] = 0;
addrs.b[z] = 0;
ovar.b[z] = 0;
}
// build list of return variables
setoutvar();
/*
* pass 1
* build aux data structure
* allocate pcs
* find use and set of variables
*/
g = flowstart(firstp, sizeof(Reg));
if(g == nil)
return;
firstr = (Reg*)g->start;
for(r = firstr; r != R; r = (Reg*)r->f.link) {
p = r->f.prog;
proginfo(&info, p);
// Avoid making variables for direct-called functions.
if(p->as == ABL && p->to.type == D_EXTERN)
continue;
bit = mkvar(r, &p->from);
if(info.flags & LeftRead)
for(z=0; z<BITS; z++)
r->use1.b[z] |= bit.b[z];
if(info.flags & LeftAddr)
setaddrs(bit);
if(info.flags & RegRead) {
if(p->from.type != D_FREG)
r->use1.b[0] |= RtoB(p->reg);
else
r->use1.b[0] |= FtoB(p->reg);
}
if(info.flags & (RightAddr | RightRead | RightWrite)) {
bit = mkvar(r, &p->to);
if(info.flags & RightAddr)
setaddrs(bit);
if(info.flags & RightRead)
for(z=0; z<BITS; z++)
r->use2.b[z] |= bit.b[z];
if(info.flags & RightWrite)
for(z=0; z<BITS; z++)
r->set.b[z] |= bit.b[z];
}
}
if(firstr == R)
return;
for(i=0; i<nvar; i++) {
Var *v = var+i;
if(v->addr) {
bit = blsh(i);
for(z=0; z<BITS; z++)
addrs.b[z] |= bit.b[z];
}
if(debug['R'] && debug['v'])
print("bit=%2d addr=%d et=%-6E w=%-2d s=%N + %lld\n",
i, v->addr, v->etype, v->width, v->node, v->offset);
}
if(debug['R'] && debug['v'])
dumpit("pass1", &firstr->f, 1);
/*
* pass 2
* find looping structure
*/
flowrpo(g);
if(debug['R'] && debug['v'])
dumpit("pass2", &firstr->f, 1);
/*
* pass 3
* iterate propagating usage
* back until flow graph is complete
*/
loop1:
change = 0;
for(r = firstr; r != R; r = (Reg*)r->f.link)
r->f.active = 0;
for(r = firstr; r != R; r = (Reg*)r->f.link)
if(r->f.prog->as == ARET)
prop(r, zbits, zbits);
loop11:
/* pick up unreachable code */
i = 0;
for(r = firstr; r != R; r = r1) {
r1 = (Reg*)r->f.link;
if(r1 && r1->f.active && !r->f.active) {
prop(r, zbits, zbits);
i = 1;
}
}
if(i)
goto loop11;
if(change)
goto loop1;
if(debug['R'] && debug['v'])
dumpit("pass3", &firstr->f, 1);
/*
* pass 4
* iterate propagating register/variable synchrony
* forward until graph is complete
*/
loop2:
change = 0;
for(r = firstr; r != R; r = (Reg*)r->f.link)
r->f.active = 0;
synch(firstr, zbits);
if(change)
goto loop2;
addsplits();
if(debug['R'] && debug['v'])
dumpit("pass4", &firstr->f, 1);
if(debug['R'] > 1) {
print("\nprop structure:\n");
for(r = firstr; r != R; r = (Reg*)r->f.link) {
print("%d:%P", r->f.loop, r->f.prog);
for(z=0; z<BITS; z++) {
bit.b[z] = r->set.b[z] |
r->refahead.b[z] | r->calahead.b[z] |
r->refbehind.b[z] | r->calbehind.b[z] |
r->use1.b[z] | r->use2.b[z];
bit.b[z] &= ~addrs.b[z];
}
if(bany(&bit)) {
print("\t");
if(bany(&r->use1))
print(" u1=%Q", r->use1);
if(bany(&r->use2))
print(" u2=%Q", r->use2);
if(bany(&r->set))
print(" st=%Q", r->set);
if(bany(&r->refahead))
print(" ra=%Q", r->refahead);
if(bany(&r->calahead))
print(" ca=%Q", r->calahead);
if(bany(&r->refbehind))
print(" rb=%Q", r->refbehind);
if(bany(&r->calbehind))
print(" cb=%Q", r->calbehind);
}
print("\n");
}
}
/*
* pass 4.5
* move register pseudo-variables into regu.
*/
for(r = firstr; r != R; r = (Reg*)r->f.link) {
r->regu = (r->refbehind.b[0] | r->set.b[0]) & REGBITS;
r->set.b[0] &= ~REGBITS;
r->use1.b[0] &= ~REGBITS;
r->use2.b[0] &= ~REGBITS;
r->refbehind.b[0] &= ~REGBITS;
r->refahead.b[0] &= ~REGBITS;
r->calbehind.b[0] &= ~REGBITS;
r->calahead.b[0] &= ~REGBITS;
r->regdiff.b[0] &= ~REGBITS;
r->act.b[0] &= ~REGBITS;
}
if(debug['R'] && debug['v'])
dumpit("pass4.5", &firstr->f, 1);
/*
* pass 5
* isolate regions
* calculate costs (paint1)
*/
r = firstr;
if(r) {
for(z=0; z<BITS; z++)
bit.b[z] = (r->refahead.b[z] | r->calahead.b[z]) &
~(externs.b[z] | params.b[z] | addrs.b[z] | consts.b[z]);
if(bany(&bit) & !r->f.refset) {
// should never happen - all variables are preset
if(debug['w'])
print("%L: used and not set: %Q\n", r->f.prog->lineno, bit);
r->f.refset = 1;
}
}
for(r = firstr; r != R; r = (Reg*)r->f.link)
r->act = zbits;
rgp = region;
nregion = 0;
for(r = firstr; r != R; r = (Reg*)r->f.link) {
for(z=0; z<BITS; z++)
bit.b[z] = r->set.b[z] &
~(r->refahead.b[z] | r->calahead.b[z] | addrs.b[z]);
if(bany(&bit) && !r->f.refset) {
if(debug['w'])
print("%L: set and not used: %Q\n", r->f.prog->lineno, bit);
r->f.refset = 1;
excise(&r->f);
}
for(z=0; z<BITS; z++)
bit.b[z] = LOAD(r) & ~(r->act.b[z] | addrs.b[z]);
while(bany(&bit)) {
i = bnum(bit);
rgp->enter = r;
rgp->varno = i;
change = 0;
if(debug['R'] > 1)
print("\n");
paint1(r, i);
bit.b[i/32] &= ~(1L<<(i%32));
if(change <= 0) {
if(debug['R'])
print("%L $%d: %Q\n",
r->f.prog->lineno, change, blsh(i));
continue;
}
rgp->cost = change;
nregion++;
if(nregion >= NRGN) {
if(debug['R'] > 1)
print("too many regions\n");
goto brk;
}
rgp++;
}
}
brk:
qsort(region, nregion, sizeof(region[0]), rcmp);
if(debug['R'] && debug['v'])
dumpit("pass5", &firstr->f, 1);
/*
* pass 6
* determine used registers (paint2)
* replace code (paint3)
*/
rgp = region;
for(i=0; i<nregion; i++) {
bit = blsh(rgp->varno);
vreg = paint2(rgp->enter, rgp->varno);
vreg = allreg(vreg, rgp);
if(debug['R']) {
if(rgp->regno >= NREG)
print("%L $%d F%d: %Q\n",
rgp->enter->f.prog->lineno,
rgp->cost,
rgp->regno-NREG,
bit);
else
print("%L $%d R%d: %Q\n",
rgp->enter->f.prog->lineno,
rgp->cost,
rgp->regno,
bit);
}
if(rgp->regno != 0)
paint3(rgp->enter, rgp->varno, vreg, rgp->regno);
rgp++;
}
if(debug['R'] && debug['v'])
dumpit("pass6", &firstr->f, 1);
/*
* pass 7
* peep-hole on basic block
*/
if(!debug['R'] || debug['P']) {
peep(firstp);
}
if(debug['R'] && debug['v'])
dumpit("pass7", &firstr->f, 1);
/*
* last pass
* eliminate nops
* free aux structures
* adjust the stack pointer
* MOVW.W R1,-12(R13) <<- start
* MOVW R0,R1
* MOVW R1,8(R13)
* MOVW $0,R1
* MOVW R1,4(R13)
* BL ,runtime.newproc+0(SB)
* MOVW &ft+-32(SP),R7 <<- adjust
* MOVW &j+-40(SP),R6 <<- adjust
* MOVW autotmp_0003+-24(SP),R5 <<- adjust
* MOVW $12(R13),R13 <<- finish
*/
vreg = 0;
for(p = firstp; p != P; p = p->link) {
while(p->link != P && p->link->as == ANOP)
p->link = p->link->link;
if(p->to.type == D_BRANCH)
while(p->to.u.branch != P && p->to.u.branch->as == ANOP)
p->to.u.branch = p->to.u.branch->link;
if(p->as == AMOVW && p->to.reg == 13) {
if(p->scond & C_WBIT) {
vreg = -p->to.offset; // in adjust region
// print("%P adjusting %d\n", p, vreg);
continue;
}
if(p->from.type == D_CONST && p->to.type == D_REG) {
if(p->from.offset != vreg)
print("in and out different\n");
// print("%P finish %d\n", p, vreg);
vreg = 0; // done adjust region
continue;
}
// print("%P %d %d from type\n", p, p->from.type, D_CONST);
// print("%P %d %d to type\n\n", p, p->to.type, D_REG);
}
if(p->as == AMOVW && vreg != 0) {
if(p->from.sym != S)
if(p->from.name == D_AUTO || p->from.name == D_PARAM) {
p->from.offset += vreg;
// print("%P adjusting from %d %d\n", p, vreg, p->from.type);
}
if(p->to.sym != S)
if(p->to.name == D_AUTO || p->to.name == D_PARAM) {
p->to.offset += vreg;
// print("%P adjusting to %d %d\n", p, vreg, p->from.type);
}
}
}
flowend(g);
}
void
regopt(Prog *p)
{
Reg *r, *r1, *r2;
Prog *p1;
int i, z;
long initpc, val, npc;
ulong vreg;
Bits bit;
struct
{
long m;
long c;
Reg* p;
} log5[6], *lp;
firstr = R;
lastr = R;
nvar = 0;
regbits = 0;
for(z=0; z<BITS; z++) {
externs.b[z] = 0;
params.b[z] = 0;
consts.b[z] = 0;
addrs.b[z] = 0;
}
/*
* pass 1
* build aux data structure
* allocate pcs
* find use and set of variables
*/
val = 5L * 5L * 5L * 5L * 5L;
lp = log5;
for(i=0; i<5; i++) {
lp->m = val;
lp->c = 0;
lp->p = R;
val /= 5L;
lp++;
}
val = 0;
for(; p != P; p = p->link) {
switch(p->as) {
case ADATA:
case AGLOBL:
case ANAME:
case ASIGNAME:
continue;
}
r = rega();
if(firstr == R) {
firstr = r;
lastr = r;
} else {
lastr->link = r;
r->p1 = lastr;
lastr->s1 = r;
lastr = r;
}
r->prog = p;
r->pc = val;
val++;
lp = log5;
for(i=0; i<5; i++) {
lp->c--;
if(lp->c <= 0) {
lp->c = lp->m;
if(lp->p != R)
lp->p->log5 = r;
lp->p = r;
(lp+1)->c = 0;
break;
}
lp++;
}
r1 = r->p1;
if(r1 != R)
switch(r1->prog->as) {
case ARETURN:
case ABR:
case ARFI:
case ARFCI:
case ARFID:
r->p1 = R;
r1->s1 = R;
}
/*
* left side always read
*/
bit = mkvar(&p->from, p->as==AMOVW || p->as == AMOVWZ || p->as == AMOVD);
for(z=0; z<BITS; z++)
r->use1.b[z] |= bit.b[z];
/*
* right side depends on opcode
*/
bit = mkvar(&p->to, 0);
if(bany(&bit))
switch(p->as) {
default:
diag(Z, "reg: unknown asop: %A", p->as);
break;
/*
* right side write
*/
case ANOP:
case AMOVB:
case AMOVBU:
case AMOVBZ:
case AMOVBZU:
case AMOVH:
case AMOVHBR:
case AMOVWBR:
case AMOVHU:
case AMOVHZ:
case AMOVHZU:
case AMOVW:
case AMOVWU:
case AMOVWZ:
case AMOVWZU:
case AMOVD:
case AMOVDU:
case AFMOVD:
case AFMOVDCC:
case AFMOVDU:
case AFMOVS:
case AFMOVSU:
case AFRSP:
for(z=0; z<BITS; z++)
r->set.b[z] |= bit.b[z];
break;
/*
* funny
*/
case ABL:
for(z=0; z<BITS; z++)
addrs.b[z] |= bit.b[z];
break;
}
}
if(firstr == R)
return;
initpc = pc - val;
npc = val;
/*
* pass 2
* turn branch references to pointers
* build back pointers
*/
for(r = firstr; r != R; r = r->link) {
p = r->prog;
if(p->to.type == D_BRANCH) {
val = p->to.offset - initpc;
r1 = firstr;
while(r1 != R) {
r2 = r1->log5;
if(r2 != R && val >= r2->pc) {
r1 = r2;
continue;
}
if(r1->pc == val)
break;
r1 = r1->link;
}
if(r1 == R) {
nearln = p->lineno;
diag(Z, "ref not found\n%P", p);
continue;
}
if(r1 == r) {
nearln = p->lineno;
diag(Z, "ref to self\n%P", p);
continue;
}
r->s2 = r1;
r->p2link = r1->p2;
r1->p2 = r;
}
}
if(debug['R']) {
p = firstr->prog;
print("\n%L %D\n", p->lineno, &p->from);
}
/*
* pass 2.5
* find looping structure
*/
for(r = firstr; r != R; r = r->link)
r->active = 0;
change = 0;
loopit(firstr, npc);
if(debug['R'] && debug['v']) {
print("\nlooping structure:\n");
for(r = firstr; r != R; r = r->link) {
print("%ld:%P", r->loop, r->prog);
for(z=0; z<BITS; z++)
bit.b[z] = r->use1.b[z] |
r->use2.b[z] | r->set.b[z];
if(bany(&bit)) {
print("\t");
if(bany(&r->use1))
print(" u1=%B", r->use1);
if(bany(&r->use2))
print(" u2=%B", r->use2);
if(bany(&r->set))
print(" st=%B", r->set);
}
print("\n");
}
}
/*
* pass 3
* iterate propagating usage
* back until flow graph is complete
*/
loop1:
change = 0;
for(r = firstr; r != R; r = r->link)
r->active = 0;
for(r = firstr; r != R; r = r->link)
if(r->prog->as == ARETURN)
prop(r, zbits, zbits);
loop11:
/* pick up unreachable code */
i = 0;
for(r = firstr; r != R; r = r1) {
r1 = r->link;
if(r1 && r1->active && !r->active) {
prop(r, zbits, zbits);
i = 1;
}
}
if(i)
goto loop11;
if(change)
goto loop1;
/*
* pass 4
* iterate propagating register/variable synchrony
* forward until graph is complete
*/
loop2:
change = 0;
for(r = firstr; r != R; r = r->link)
r->active = 0;
synch(firstr, zbits);
if(change)
goto loop2;
/*
* pass 5
* isolate regions
* calculate costs (paint1)
*/
r = firstr;
if(r) {
for(z=0; z<BITS; z++)
bit.b[z] = (r->refahead.b[z] | r->calahead.b[z]) &
~(externs.b[z] | params.b[z] | addrs.b[z] | consts.b[z]);
if(bany(&bit)) {
nearln = r->prog->lineno;
warn(Z, "used and not set: %B", bit);
if(debug['R'] && !debug['w'])
print("used and not set: %B\n", bit);
}
}
if(debug['R'] && debug['v'])
print("\nprop structure:\n");
for(r = firstr; r != R; r = r->link)
r->act = zbits;
rgp = region;
nregion = 0;
for(r = firstr; r != R; r = r->link) {
if(debug['R'] && debug['v'])
print("%P\n set = %B; rah = %B; cal = %B\n",
r->prog, r->set, r->refahead, r->calahead);
for(z=0; z<BITS; z++)
bit.b[z] = r->set.b[z] &
~(r->refahead.b[z] | r->calahead.b[z] | addrs.b[z]);
if(bany(&bit)) {
nearln = r->prog->lineno;
warn(Z, "set and not used: %B", bit);
if(debug['R'])
print("set an not used: %B\n", bit);
excise(r);
}
for(z=0; z<BITS; z++)
bit.b[z] = LOAD(r) & ~(r->act.b[z] | addrs.b[z]);
while(bany(&bit)) {
i = bnum(bit);
rgp->enter = r;
rgp->varno = i;
change = 0;
if(debug['R'] && debug['v'])
print("\n");
paint1(r, i);
bit.b[i/32] &= ~(1L<<(i%32));
if(change <= 0) {
if(debug['R'])
print("%L$%d: %B\n",
r->prog->lineno, change, blsh(i));
continue;
}
rgp->cost = change;
nregion++;
if(nregion >= NRGN) {
warn(Z, "too many regions");
goto brk;
}
rgp++;
}
}
brk:
qsort(region, nregion, sizeof(region[0]), rcmp);
/*
* pass 6
* determine used registers (paint2)
* replace code (paint3)
*/
rgp = region;
for(i=0; i<nregion; i++) {
bit = blsh(rgp->varno);
vreg = paint2(rgp->enter, rgp->varno);
vreg = allreg(vreg, rgp);
if(debug['R']) {
if(rgp->regno >= NREG)
print("%L$%d F%d: %B\n",
rgp->enter->prog->lineno,
rgp->cost,
rgp->regno-NREG,
bit);
else
print("%L$%d R%d: %B\n",
rgp->enter->prog->lineno,
rgp->cost,
rgp->regno,
bit);
}
if(rgp->regno != 0)
paint3(rgp->enter, rgp->varno, vreg, rgp->regno);
rgp++;
}
/*
* pass 7
* peep-hole on basic block
*/
if(!debug['R'] || debug['P'])
peep();
/*
* pass 8
* recalculate pc
*/
val = initpc;
for(r = firstr; r != R; r = r1) {
r->pc = val;
p = r->prog;
p1 = P;
r1 = r->link;
if(r1 != R)
p1 = r1->prog;
for(; p != p1; p = p->link) {
switch(p->as) {
default:
val++;
break;
case ANOP:
case ADATA:
case AGLOBL:
case ANAME:
case ASIGNAME:
break;
}
}
}
pc = val;
/*
* fix up branches
*/
if(debug['R'])
if(bany(&addrs))
print("addrs: %B\n", addrs);
r1 = 0; /* set */
for(r = firstr; r != R; r = r->link) {
p = r->prog;
if(p->to.type == D_BRANCH)
p->to.offset = r->s2->pc;
r1 = r;
}
/*
* last pass
* eliminate nops
* free aux structures
*/
for(p = firstr->prog; p != P; p = p->link){
while(p->link && p->link->as == ANOP)
p->link = p->link->link;
}
if(r1 != R) {
r1->link = freer;
freer = firstr;
}
}
void
regopt(Prog *p)
{
Reg *r, *r1, *r2;
Prog *p1;
int i, z;
int32_t initpc, val, npc;
uint32_t vreg;
Bits bit;
struct
{
int32_t m;
int32_t c;
Reg* p;
} log5[6], *lp;
firstr = R;
lastr = R;
nvar = 0;
regbits = RtoB(D_SP) | RtoB(D_AX);
for(z=0; z<BITS; z++) {
externs.b[z] = 0;
params.b[z] = 0;
consts.b[z] = 0;
addrs.b[z] = 0;
}
/*
* pass 1
* build aux data structure
* allocate pcs
* find use and set of variables
*/
val = 5L * 5L * 5L * 5L * 5L;
lp = log5;
for(i=0; i<5; i++) {
lp->m = val;
lp->c = 0;
lp->p = R;
val /= 5L;
lp++;
}
val = 0;
for(; p != P; p = p->link) {
switch(p->as) {
case ADATA:
case AGLOBL:
case ANAME:
case ASIGNAME:
continue;
}
r = rega();
if(firstr == R) {
firstr = r;
lastr = r;
} else {
lastr->link = r;
r->p1 = lastr;
lastr->s1 = r;
lastr = r;
}
r->prog = p;
r->pc = val;
val++;
lp = log5;
for(i=0; i<5; i++) {
lp->c--;
if(lp->c <= 0) {
lp->c = lp->m;
if(lp->p != R)
lp->p->log5 = r;
lp->p = r;
(lp+1)->c = 0;
break;
}
lp++;
}
r1 = r->p1;
if(r1 != R)
switch(r1->prog->as) {
case ARET:
case AJMP:
case AIRETL:
r->p1 = R;
r1->s1 = R;
}
bit = mkvar(r, &p->from, p->as==AMOVL);
if(bany(&bit))
switch(p->as) {
/*
* funny
*/
case ALEAL:
for(z=0; z<BITS; z++)
addrs.b[z] |= bit.b[z];
break;
/*
* left side read
*/
default:
for(z=0; z<BITS; z++)
r->use1.b[z] |= bit.b[z];
break;
}
bit = mkvar(r, &p->to, 0);
if(bany(&bit))
switch(p->as) {
default:
diag(Z, "reg: unknown op: %A", p->as);
break;
/*
* right side read
*/
case ACMPB:
case ACMPL:
case ACMPW:
for(z=0; z<BITS; z++)
r->use2.b[z] |= bit.b[z];
break;
/*
* right side write
*/
case ANOP:
case AMOVL:
case AMOVB:
case AMOVW:
case AMOVBLSX:
case AMOVBLZX:
case AMOVWLSX:
case AMOVWLZX:
for(z=0; z<BITS; z++)
r->set.b[z] |= bit.b[z];
break;
/*
* right side read+write
*/
case AADDB:
case AADDL:
case AADDW:
case AANDB:
case AANDL:
case AANDW:
case ASUBB:
case ASUBL:
case ASUBW:
case AORB:
case AORL:
case AORW:
case AXORB:
case AXORL:
case AXORW:
case ASALB:
case ASALL:
case ASALW:
case ASARB:
case ASARL:
case ASARW:
case AROLB:
case AROLL:
case AROLW:
case ARORB:
case ARORL:
case ARORW:
case ASHLB:
case ASHLL:
case ASHLW:
case ASHRB:
case ASHRL:
case ASHRW:
case AIMULL:
case AIMULW:
case ANEGL:
case ANOTL:
case AADCL:
case ASBBL:
for(z=0; z<BITS; z++) {
r->set.b[z] |= bit.b[z];
r->use2.b[z] |= bit.b[z];
}
break;
/*
* funny
*/
case AFMOVDP:
case AFMOVFP:
case AFMOVLP:
case AFMOVVP:
case AFMOVWP:
case ACALL:
for(z=0; z<BITS; z++)
addrs.b[z] |= bit.b[z];
break;
}
switch(p->as) {
case AIMULL:
case AIMULW:
if(p->to.type != D_NONE)
break;
case AIDIVB:
case AIDIVL:
case AIDIVW:
case AIMULB:
case ADIVB:
case ADIVL:
case ADIVW:
case AMULB:
case AMULL:
case AMULW:
case ACWD:
case ACDQ:
r->regu |= RtoB(D_AX) | RtoB(D_DX);
break;
case AREP:
case AREPN:
case ALOOP:
case ALOOPEQ:
case ALOOPNE:
r->regu |= RtoB(D_CX);
break;
case AMOVSB:
case AMOVSL:
case AMOVSW:
case ACMPSB:
case ACMPSL:
case ACMPSW:
r->regu |= RtoB(D_SI) | RtoB(D_DI);
break;
case ASTOSB:
case ASTOSL:
case ASTOSW:
case ASCASB:
case ASCASL:
case ASCASW:
r->regu |= RtoB(D_AX) | RtoB(D_DI);
break;
case AINSB:
case AINSL:
case AINSW:
case AOUTSB:
case AOUTSL:
case AOUTSW:
r->regu |= RtoB(D_DI) | RtoB(D_DX);
break;
case AFSTSW:
case ASAHF:
r->regu |= RtoB(D_AX);
break;
}
}
if(firstr == R)
return;
initpc = pc - val;
npc = val;
/*
* pass 2
* turn branch references to pointers
* build back pointers
*/
for(r = firstr; r != R; r = r->link) {
p = r->prog;
if(p->to.type == D_BRANCH) {
val = p->to.offset - initpc;
r1 = firstr;
while(r1 != R) {
r2 = r1->log5;
if(r2 != R && val >= r2->pc) {
r1 = r2;
continue;
}
if(r1->pc == val)
break;
r1 = r1->link;
}
if(r1 == R) {
nearln = p->lineno;
diag(Z, "ref not found\n%P", p);
continue;
}
if(r1 == r) {
nearln = p->lineno;
diag(Z, "ref to self\n%P", p);
continue;
}
r->s2 = r1;
r->p2link = r1->p2;
r1->p2 = r;
}
}
if(debug['R']) {
p = firstr->prog;
print("\n%L %D\n", p->lineno, &p->from);
}
/*
* pass 2.5
* find looping structure
*/
for(r = firstr; r != R; r = r->link)
r->active = 0;
change = 0;
loopit(firstr, npc);
if(debug['R'] && debug['v']) {
print("\nlooping structure:\n");
for(r = firstr; r != R; r = r->link) {
print("%ld:%P", r->loop, r->prog);
for(z=0; z<BITS; z++)
bit.b[z] = r->use1.b[z] |
r->use2.b[z] |
r->set.b[z];
if(bany(&bit)) {
print("\t");
if(bany(&r->use1))
print(" u1=%B", r->use1);
if(bany(&r->use2))
print(" u2=%B", r->use2);
if(bany(&r->set))
print(" st=%B", r->set);
}
print("\n");
}
}
/*
* pass 3
* iterate propagating usage
* back until flow graph is complete
*/
loop1:
change = 0;
for(r = firstr; r != R; r = r->link)
r->active = 0;
for(r = firstr; r != R; r = r->link)
if(r->prog->as == ARET)
prop(r, zbits, zbits);
loop11:
/* pick up unreachable code */
i = 0;
for(r = firstr; r != R; r = r1) {
r1 = r->link;
if(r1 && r1->active && !r->active) {
prop(r, zbits, zbits);
i = 1;
}
}
if(i)
goto loop11;
if(change)
goto loop1;
/*
* pass 4
* iterate propagating register/variable synchrony
* forward until graph is complete
*/
loop2:
change = 0;
for(r = firstr; r != R; r = r->link)
r->active = 0;
synch(firstr, zbits);
if(change)
goto loop2;
/*
* pass 5
* isolate regions
* calculate costs (paint1)
*/
r = firstr;
if(r) {
for(z=0; z<BITS; z++)
bit.b[z] = (r->refahead.b[z] | r->calahead.b[z]) &
~(externs.b[z] | params.b[z] | addrs.b[z] | consts.b[z]);
if(bany(&bit)) {
nearln = r->prog->lineno;
warn(Z, "used and not set: %B", bit);
if(debug['R'] && !debug['w'])
print("used and not set: %B\n", bit);
}
}
if(debug['R'] && debug['v'])
print("\nprop structure:\n");
for(r = firstr; r != R; r = r->link)
r->act = zbits;
rgp = region;
nregion = 0;
for(r = firstr; r != R; r = r->link) {
if(debug['R'] && debug['v']) {
print("%P\t", r->prog);
if(bany(&r->set))
print("s:%B ", r->set);
if(bany(&r->refahead))
print("ra:%B ", r->refahead);
if(bany(&r->calahead))
print("ca:%B ", r->calahead);
print("\n");
}
for(z=0; z<BITS; z++)
bit.b[z] = r->set.b[z] &
~(r->refahead.b[z] | r->calahead.b[z] | addrs.b[z]);
if(bany(&bit)) {
nearln = r->prog->lineno;
warn(Z, "set and not used: %B", bit);
if(debug['R'])
print("set and not used: %B\n", bit);
excise(r);
}
for(z=0; z<BITS; z++)
bit.b[z] = LOAD(r) & ~(r->act.b[z] | addrs.b[z]);
while(bany(&bit)) {
i = bnum(bit);
rgp->enter = r;
rgp->varno = i;
change = 0;
if(debug['R'] && debug['v'])
print("\n");
paint1(r, i);
bit.b[i/32] &= ~(1L<<(i%32));
if(change <= 0) {
if(debug['R'])
print("%L$%d: %B\n",
r->prog->lineno, change, blsh(i));
continue;
}
rgp->cost = change;
nregion++;
if(nregion >= NRGN) {
warn(Z, "too many regions");
goto brk;
}
rgp++;
}
}
brk:
qsort(region, nregion, sizeof(region[0]), rcmp);
/*
* pass 6
* determine used registers (paint2)
* replace code (paint3)
*/
rgp = region;
for(i=0; i<nregion; i++) {
bit = blsh(rgp->varno);
vreg = paint2(rgp->enter, rgp->varno);
vreg = allreg(vreg, rgp);
if(debug['R']) {
print("%L$%d %R: %B\n",
rgp->enter->prog->lineno,
rgp->cost,
rgp->regno,
bit);
}
if(rgp->regno != 0)
paint3(rgp->enter, rgp->varno, vreg, rgp->regno);
rgp++;
}
/*
* pass 7
* peep-hole on basic block
*/
if(!debug['R'] || debug['P'])
peep();
/*
* pass 8
* recalculate pc
*/
val = initpc;
for(r = firstr; r != R; r = r1) {
r->pc = val;
p = r->prog;
p1 = P;
r1 = r->link;
if(r1 != R)
p1 = r1->prog;
for(; p != p1; p = p->link) {
switch(p->as) {
default:
val++;
break;
case ANOP:
case ADATA:
case AGLOBL:
case ANAME:
case ASIGNAME:
break;
}
}
}
pc = val;
/*
* fix up branches
*/
if(debug['R'])
if(bany(&addrs))
print("addrs: %B\n", addrs);
r1 = 0; /* set */
for(r = firstr; r != R; r = r->link) {
p = r->prog;
if(p->to.type == D_BRANCH)
p->to.offset = r->s2->pc;
r1 = r;
}
/*
* last pass
* eliminate nops
* free aux structures
*/
for(p = firstr->prog; p != P; p = p->link){
while(p->link && p->link->as == ANOP)
p->link = p->link->link;
}
if(r1 != R) {
r1->link = freer;
freer = firstr;
}
}
Bits
mkvar(Reg *r, Adr *a)
{
Var *v;
int i, t, n, et, z, w, flag;
uint32 regu;
int32 o;
Bits bit;
Sym *s;
/*
* mark registers used
*/
t = a->type;
if(t == D_NONE)
goto none;
if(r != R)
r->use1.b[0] |= doregbits(a->index);
switch(t) {
default:
regu = doregbits(t);
if(regu == 0)
goto none;
bit = zbits;
bit.b[0] = regu;
return bit;
case D_ADDR:
a->type = a->index;
bit = mkvar(r, a);
setaddrs(bit);
a->type = t;
ostats.naddr++;
goto none;
case D_EXTERN:
case D_STATIC:
case D_PARAM:
case D_AUTO:
n = t;
break;
}
s = a->sym;
if(s == S)
goto none;
if(s->name[0] == '.')
goto none;
et = a->etype;
o = a->offset;
w = a->width;
flag = 0;
for(i=0; i<nvar; i++) {
v = var+i;
if(v->sym == s && v->name == n) {
if(v->offset == o)
if(v->etype == et)
if(v->width == w)
return blsh(i);
// if they overlaps, disable both
if(overlap(v->offset, v->width, o, w)) {
// print("disable overlap %s %d %d %d %d, %E != %E\n", s->name, v->offset, v->width, o, w, v->etype, et);
v->addr = 1;
flag = 1;
}
}
}
if(a->pun) {
// print("disable pun %s\n", s->name);
flag = 1;
}
switch(et) {
case 0:
case TFUNC:
goto none;
}
if(nvar >= NVAR) {
if(debug['w'] > 1 && s)
fatal("variable not optimized: %D", a);
goto none;
}
i = nvar;
nvar++;
v = var+i;
v->sym = s;
v->offset = o;
v->name = n;
v->gotype = a->gotype;
v->etype = et;
v->width = w;
v->addr = flag; // funny punning
if(debug['R'])
print("bit=%2d et=%2d w=%d %S %D\n", i, et, w, s, a);
ostats.nvar++;
bit = blsh(i);
if(n == D_EXTERN || n == D_STATIC)
for(z=0; z<BITS; z++)
externs.b[z] |= bit.b[z];
if(n == D_PARAM)
for(z=0; z<BITS; z++)
params.b[z] |= bit.b[z];
return bit;
none:
return zbits;
}
void
regopt(Prog *firstp)
{
Reg *r, *r1;
Prog *p;
Graph *g;
ProgInfo info;
int i, z;
uint32 vreg;
Bits bit;
if(first) {
fmtinstall('Q', Qconv);
exregoffset = D_R15;
first = 0;
}
mergetemp(firstp);
/*
* control flow is more complicated in generated go code
* than in generated c code. define pseudo-variables for
* registers, so we have complete register usage information.
*/
nvar = NREGVAR;
memset(var, 0, NREGVAR*sizeof var[0]);
for(i=0; i<NREGVAR; i++) {
if(regnodes[i] == N)
regnodes[i] = newname(lookup(regname[i]));
var[i].node = regnodes[i];
}
regbits = RtoB(D_SP);
for(z=0; z<BITS; z++) {
externs.b[z] = 0;
params.b[z] = 0;
consts.b[z] = 0;
addrs.b[z] = 0;
ovar.b[z] = 0;
}
// build list of return variables
setoutvar();
/*
* pass 1
* build aux data structure
* allocate pcs
* find use and set of variables
*/
g = flowstart(firstp, sizeof(Reg));
if(g == nil)
return;
firstr = (Reg*)g->start;
for(r = firstr; r != R; r = (Reg*)r->f.link) {
p = r->f.prog;
if(p->as == AVARDEF)
continue;
proginfo(&info, p);
// Avoid making variables for direct-called functions.
if(p->as == ACALL && p->to.type == D_EXTERN)
continue;
r->use1.b[0] |= info.reguse | info.regindex;
r->set.b[0] |= info.regset;
bit = mkvar(r, &p->from);
if(bany(&bit)) {
if(info.flags & LeftAddr)
setaddrs(bit);
if(info.flags & LeftRead)
for(z=0; z<BITS; z++)
r->use1.b[z] |= bit.b[z];
if(info.flags & LeftWrite)
for(z=0; z<BITS; z++)
r->set.b[z] |= bit.b[z];
}
bit = mkvar(r, &p->to);
if(bany(&bit)) {
if(info.flags & RightAddr)
setaddrs(bit);
if(info.flags & RightRead)
for(z=0; z<BITS; z++)
r->use2.b[z] |= bit.b[z];
if(info.flags & RightWrite)
for(z=0; z<BITS; z++)
r->set.b[z] |= bit.b[z];
}
}
for(i=0; i<nvar; i++) {
Var *v = var+i;
if(v->addr) {
bit = blsh(i);
for(z=0; z<BITS; z++)
addrs.b[z] |= bit.b[z];
}
if(debug['R'] && debug['v'])
print("bit=%2d addr=%d et=%-6E w=%-2d s=%N + %lld\n",
i, v->addr, v->etype, v->width, v->node, v->offset);
}
if(debug['R'] && debug['v'])
dumpit("pass1", &firstr->f, 1);
/*
* pass 2
* find looping structure
*/
flowrpo(g);
if(debug['R'] && debug['v'])
dumpit("pass2", &firstr->f, 1);
/*
* pass 3
* iterate propagating usage
* back until flow graph is complete
*/
loop1:
change = 0;
for(r = firstr; r != R; r = (Reg*)r->f.link)
r->f.active = 0;
for(r = firstr; r != R; r = (Reg*)r->f.link)
if(r->f.prog->as == ARET)
prop(r, zbits, zbits);
loop11:
/* pick up unreachable code */
i = 0;
for(r = firstr; r != R; r = r1) {
r1 = (Reg*)r->f.link;
if(r1 && r1->f.active && !r->f.active) {
prop(r, zbits, zbits);
i = 1;
}
}
if(i)
goto loop11;
if(change)
goto loop1;
if(debug['R'] && debug['v'])
dumpit("pass3", &firstr->f, 1);
/*
* pass 4
* iterate propagating register/variable synchrony
* forward until graph is complete
*/
loop2:
change = 0;
for(r = firstr; r != R; r = (Reg*)r->f.link)
r->f.active = 0;
synch(firstr, zbits);
if(change)
goto loop2;
if(debug['R'] && debug['v'])
dumpit("pass4", &firstr->f, 1);
/*
* pass 4.5
* move register pseudo-variables into regu.
*/
for(r = firstr; r != R; r = (Reg*)r->f.link) {
r->regu = (r->refbehind.b[0] | r->set.b[0]) & REGBITS;
r->set.b[0] &= ~REGBITS;
r->use1.b[0] &= ~REGBITS;
r->use2.b[0] &= ~REGBITS;
r->refbehind.b[0] &= ~REGBITS;
r->refahead.b[0] &= ~REGBITS;
r->calbehind.b[0] &= ~REGBITS;
r->calahead.b[0] &= ~REGBITS;
r->regdiff.b[0] &= ~REGBITS;
r->act.b[0] &= ~REGBITS;
}
/*
* pass 5
* isolate regions
* calculate costs (paint1)
*/
r = firstr;
if(r) {
for(z=0; z<BITS; z++)
bit.b[z] = (r->refahead.b[z] | r->calahead.b[z]) &
~(externs.b[z] | params.b[z] | addrs.b[z] | consts.b[z]);
if(bany(&bit) && !r->f.refset) {
// should never happen - all variables are preset
if(debug['w'])
print("%L: used and not set: %Q\n", r->f.prog->lineno, bit);
r->f.refset = 1;
}
}
for(r = firstr; r != R; r = (Reg*)r->f.link)
r->act = zbits;
rgp = region;
nregion = 0;
for(r = firstr; r != R; r = (Reg*)r->f.link) {
for(z=0; z<BITS; z++)
bit.b[z] = r->set.b[z] &
~(r->refahead.b[z] | r->calahead.b[z] | addrs.b[z]);
if(bany(&bit) && !r->f.refset) {
if(debug['w'])
print("%L: set and not used: %Q\n", r->f.prog->lineno, bit);
r->f.refset = 1;
excise(&r->f);
}
for(z=0; z<BITS; z++)
bit.b[z] = LOAD(r) & ~(r->act.b[z] | addrs.b[z]);
while(bany(&bit)) {
i = bnum(bit);
rgp->enter = r;
rgp->varno = i;
change = 0;
paint1(r, i);
bit.b[i/32] &= ~(1L<<(i%32));
if(change <= 0)
continue;
rgp->cost = change;
nregion++;
if(nregion >= NRGN) {
if(debug['R'] && debug['v'])
print("too many regions\n");
goto brk;
}
rgp++;
}
}
brk:
qsort(region, nregion, sizeof(region[0]), rcmp);
if(debug['R'] && debug['v'])
dumpit("pass5", &firstr->f, 1);
/*
* pass 6
* determine used registers (paint2)
* replace code (paint3)
*/
rgp = region;
for(i=0; i<nregion; i++) {
bit = blsh(rgp->varno);
vreg = paint2(rgp->enter, rgp->varno);
vreg = allreg(vreg, rgp);
if(rgp->regno != 0) {
if(debug['R'] && debug['v']) {
Var *v;
v = var + rgp->varno;
print("registerize %N+%lld (bit=%2d et=%2E) in %R\n",
v->node, v->offset, rgp->varno, v->etype, rgp->regno);
}
paint3(rgp->enter, rgp->varno, vreg, rgp->regno);
}
rgp++;
}
if(debug['R'] && debug['v'])
dumpit("pass6", &firstr->f, 1);
/*
* free aux structures. peep allocates new ones.
*/
flowend(g);
firstr = R;
/*
* pass 7
* peep-hole on basic block
*/
if(!debug['R'] || debug['P'])
peep(firstp);
/*
* eliminate nops
*/
for(p=firstp; p!=P; p=p->link) {
while(p->link != P && p->link->as == ANOP)
p->link = p->link->link;
if(p->to.type == D_BRANCH)
while(p->to.u.branch != P && p->to.u.branch->as == ANOP)
p->to.u.branch = p->to.u.branch->link;
}
if(debug['R']) {
if(ostats.ncvtreg ||
ostats.nspill ||
ostats.nreload ||
ostats.ndelmov ||
ostats.nvar ||
ostats.naddr ||
0)
print("\nstats\n");
if(ostats.ncvtreg)
print(" %4d cvtreg\n", ostats.ncvtreg);
if(ostats.nspill)
print(" %4d spill\n", ostats.nspill);
if(ostats.nreload)
print(" %4d reload\n", ostats.nreload);
if(ostats.ndelmov)
print(" %4d delmov\n", ostats.ndelmov);
if(ostats.nvar)
print(" %4d var\n", ostats.nvar);
if(ostats.naddr)
print(" %4d addr\n", ostats.naddr);
memset(&ostats, 0, sizeof(ostats));
}
}
void
regopt(Prog *firstp)
{
Reg *r, *r1;
Prog *p;
int i, z, nr;
uint32 vreg;
Bits bit;
if(first) {
fmtinstall('Q', Qconv);
exregoffset = D_DI; // no externals
first = 0;
}
fixjmp(firstp);
// count instructions
nr = 0;
for(p=firstp; p!=P; p=p->link)
nr++;
// if too big dont bother
if(nr >= 10000) {
// print("********** %S is too big (%d)\n", curfn->nname->sym, nr);
return;
}
r1 = R;
firstr = R;
lastr = R;
/*
* control flow is more complicated in generated go code
* than in generated c code. define pseudo-variables for
* registers, so we have complete register usage information.
*/
nvar = NREGVAR;
memset(var, 0, NREGVAR*sizeof var[0]);
for(i=0; i<NREGVAR; i++)
var[i].node = newname(lookup(regname[i]));
regbits = RtoB(D_SP);
for(z=0; z<BITS; z++) {
externs.b[z] = 0;
params.b[z] = 0;
consts.b[z] = 0;
addrs.b[z] = 0;
ovar.b[z] = 0;
}
// build list of return variables
setoutvar();
/*
* pass 1
* build aux data structure
* allocate pcs
* find use and set of variables
*/
nr = 0;
for(p=firstp; p!=P; p=p->link) {
switch(p->as) {
case ADATA:
case AGLOBL:
case ANAME:
case ASIGNAME:
continue;
}
r = rega();
nr++;
if(firstr == R) {
firstr = r;
lastr = r;
} else {
lastr->link = r;
r->p1 = lastr;
lastr->s1 = r;
lastr = r;
}
r->prog = p;
p->reg = r;
r1 = r->p1;
if(r1 != R) {
switch(r1->prog->as) {
case ARET:
case AJMP:
case AIRETL:
r->p1 = R;
r1->s1 = R;
}
}
bit = mkvar(r, &p->from);
if(bany(&bit))
switch(p->as) {
/*
* funny
*/
case ALEAL:
case AFMOVL:
case AFMOVW:
case AFMOVV:
setaddrs(bit);
break;
/*
* left side read
*/
default:
for(z=0; z<BITS; z++)
r->use1.b[z] |= bit.b[z];
break;
/*
* left side read+write
*/
case AXCHGB:
case AXCHGW:
case AXCHGL:
for(z=0; z<BITS; z++) {
r->use1.b[z] |= bit.b[z];
r->set.b[z] |= bit.b[z];
}
break;
}
bit = mkvar(r, &p->to);
if(bany(&bit))
switch(p->as) {
default:
yyerror("reg: unknown op: %A", p->as);
break;
/*
* right side read
*/
case ACMPB:
case ACMPL:
case ACMPW:
case ATESTB:
case ATESTL:
case ATESTW:
for(z=0; z<BITS; z++)
r->use2.b[z] |= bit.b[z];
break;
/*
* right side write
*/
case AFSTSW:
case ALEAL:
case ANOP:
case AMOVL:
case AMOVB:
case AMOVW:
case AMOVBLSX:
case AMOVBLZX:
case AMOVBWSX:
case AMOVBWZX:
case AMOVWLSX:
case AMOVWLZX:
case APOPL:
for(z=0; z<BITS; z++)
r->set.b[z] |= bit.b[z];
break;
/*
* right side read+write
*/
case AINCB:
case AINCL:
case AINCW:
case ADECB:
case ADECL:
case ADECW:
case AADDB:
case AADDL:
case AADDW:
case AANDB:
case AANDL:
case AANDW:
case ASUBB:
case ASUBL:
case ASUBW:
case AORB:
case AORL:
case AORW:
case AXORB:
case AXORL:
case AXORW:
case ASALB:
case ASALL:
case ASALW:
case ASARB:
case ASARL:
case ASARW:
case ARCLB:
case ARCLL:
case ARCLW:
case ARCRB:
case ARCRL:
case ARCRW:
case AROLB:
case AROLL:
case AROLW:
case ARORB:
case ARORL:
case ARORW:
case ASHLB:
case ASHLL:
case ASHLW:
case ASHRB:
case ASHRL:
case ASHRW:
case AIMULL:
case AIMULW:
case ANEGB:
case ANEGL:
case ANEGW:
case ANOTB:
case ANOTL:
case ANOTW:
case AADCL:
case ASBBL:
case ASETCC:
case ASETCS:
case ASETEQ:
case ASETGE:
case ASETGT:
case ASETHI:
case ASETLE:
case ASETLS:
case ASETLT:
case ASETMI:
case ASETNE:
case ASETOC:
case ASETOS:
case ASETPC:
case ASETPL:
case ASETPS:
case AXCHGB:
case AXCHGW:
case AXCHGL:
for(z=0; z<BITS; z++) {
r->set.b[z] |= bit.b[z];
r->use2.b[z] |= bit.b[z];
}
break;
/*
* funny
*/
case AFMOVDP:
case AFMOVFP:
case AFMOVLP:
case AFMOVVP:
case AFMOVWP:
case ACALL:
setaddrs(bit);
break;
}
switch(p->as) {
case AIMULL:
case AIMULW:
if(p->to.type != D_NONE)
break;
case AIDIVL:
case AIDIVW:
case ADIVL:
case ADIVW:
case AMULL:
case AMULW:
r->set.b[0] |= RtoB(D_AX) | RtoB(D_DX);
r->use1.b[0] |= RtoB(D_AX) | RtoB(D_DX);
break;
case AIDIVB:
case AIMULB:
case ADIVB:
case AMULB:
r->set.b[0] |= RtoB(D_AX);
r->use1.b[0] |= RtoB(D_AX);
break;
case ACWD:
r->set.b[0] |= RtoB(D_AX) | RtoB(D_DX);
r->use1.b[0] |= RtoB(D_AX);
break;
case ACDQ:
r->set.b[0] |= RtoB(D_DX);
r->use1.b[0] |= RtoB(D_AX);
break;
case AREP:
case AREPN:
case ALOOP:
case ALOOPEQ:
case ALOOPNE:
r->set.b[0] |= RtoB(D_CX);
r->use1.b[0] |= RtoB(D_CX);
break;
case AMOVSB:
case AMOVSL:
case AMOVSW:
case ACMPSB:
case ACMPSL:
case ACMPSW:
r->set.b[0] |= RtoB(D_SI) | RtoB(D_DI);
r->use1.b[0] |= RtoB(D_SI) | RtoB(D_DI);
break;
case ASTOSB:
case ASTOSL:
case ASTOSW:
case ASCASB:
case ASCASL:
case ASCASW:
r->set.b[0] |= RtoB(D_DI);
r->use1.b[0] |= RtoB(D_AX) | RtoB(D_DI);
break;
case AINSB:
case AINSL:
case AINSW:
r->set.b[0] |= RtoB(D_DX) | RtoB(D_DI);
r->use1.b[0] |= RtoB(D_DI);
break;
case AOUTSB:
case AOUTSL:
case AOUTSW:
r->set.b[0] |= RtoB(D_DI);
r->use1.b[0] |= RtoB(D_DX) | RtoB(D_DI);
break;
}
}
if(firstr == R)
return;
for(i=0; i<nvar; i++) {
Var *v = var+i;
if(v->addr) {
bit = blsh(i);
for(z=0; z<BITS; z++)
addrs.b[z] |= bit.b[z];
}
// print("bit=%2d addr=%d et=%-6E w=%-2d s=%S + %lld\n",
// i, v->addr, v->etype, v->width, v->sym, v->offset);
}
if(debug['R'] && debug['v'])
dumpit("pass1", firstr);
/*
* pass 2
* turn branch references to pointers
* build back pointers
*/
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
if(p->to.type == D_BRANCH) {
if(p->to.branch == P)
fatal("pnil %P", p);
r1 = p->to.branch->reg;
if(r1 == R)
fatal("rnil %P", p);
if(r1 == r) {
//fatal("ref to self %P", p);
continue;
}
r->s2 = r1;
r->p2link = r1->p2;
r1->p2 = r;
}
}
if(debug['R'] && debug['v'])
dumpit("pass2", firstr);
/*
* pass 2.5
* find looping structure
*/
for(r = firstr; r != R; r = r->link)
r->active = 0;
change = 0;
loopit(firstr, nr);
if(debug['R'] && debug['v'])
dumpit("pass2.5", firstr);
/*
* pass 3
* iterate propagating usage
* back until flow graph is complete
*/
loop1:
change = 0;
for(r = firstr; r != R; r = r->link)
r->active = 0;
for(r = firstr; r != R; r = r->link)
if(r->prog->as == ARET)
prop(r, zbits, zbits);
loop11:
/* pick up unreachable code */
i = 0;
for(r = firstr; r != R; r = r1) {
r1 = r->link;
if(r1 && r1->active && !r->active) {
prop(r, zbits, zbits);
i = 1;
}
}
if(i)
goto loop11;
if(change)
goto loop1;
if(debug['R'] && debug['v'])
dumpit("pass3", firstr);
/*
* pass 4
* iterate propagating register/variable synchrony
* forward until graph is complete
*/
loop2:
change = 0;
for(r = firstr; r != R; r = r->link)
r->active = 0;
synch(firstr, zbits);
if(change)
goto loop2;
if(debug['R'] && debug['v'])
dumpit("pass4", firstr);
/*
* pass 4.5
* move register pseudo-variables into regu.
*/
for(r = firstr; r != R; r = r->link) {
r->regu = (r->refbehind.b[0] | r->set.b[0]) & REGBITS;
r->set.b[0] &= ~REGBITS;
r->use1.b[0] &= ~REGBITS;
r->use2.b[0] &= ~REGBITS;
r->refbehind.b[0] &= ~REGBITS;
r->refahead.b[0] &= ~REGBITS;
r->calbehind.b[0] &= ~REGBITS;
r->calahead.b[0] &= ~REGBITS;
r->regdiff.b[0] &= ~REGBITS;
r->act.b[0] &= ~REGBITS;
}
/*
* pass 5
* isolate regions
* calculate costs (paint1)
*/
r = firstr;
if(r) {
for(z=0; z<BITS; z++)
bit.b[z] = (r->refahead.b[z] | r->calahead.b[z]) &
~(externs.b[z] | params.b[z] | addrs.b[z] | consts.b[z]);
if(bany(&bit) && !r->refset) {
// should never happen - all variables are preset
if(debug['w'])
print("%L: used and not set: %Q\n", r->prog->lineno, bit);
r->refset = 1;
}
}
for(r = firstr; r != R; r = r->link)
r->act = zbits;
rgp = region;
nregion = 0;
for(r = firstr; r != R; r = r->link) {
for(z=0; z<BITS; z++)
bit.b[z] = r->set.b[z] &
~(r->refahead.b[z] | r->calahead.b[z] | addrs.b[z]);
if(bany(&bit) && !r->refset) {
if(debug['w'])
print("%L: set and not used: %Q\n", r->prog->lineno, bit);
r->refset = 1;
excise(r);
}
for(z=0; z<BITS; z++)
bit.b[z] = LOAD(r) & ~(r->act.b[z] | addrs.b[z]);
while(bany(&bit)) {
i = bnum(bit);
rgp->enter = r;
rgp->varno = i;
change = 0;
paint1(r, i);
bit.b[i/32] &= ~(1L<<(i%32));
if(change <= 0)
continue;
rgp->cost = change;
nregion++;
if(nregion >= NRGN) {
if(debug['R'] && debug['v'])
print("too many regions\n");
goto brk;
}
rgp++;
}
}
brk:
qsort(region, nregion, sizeof(region[0]), rcmp);
/*
* pass 6
* determine used registers (paint2)
* replace code (paint3)
*/
rgp = region;
for(i=0; i<nregion; i++) {
bit = blsh(rgp->varno);
vreg = paint2(rgp->enter, rgp->varno);
vreg = allreg(vreg, rgp);
if(rgp->regno != 0)
paint3(rgp->enter, rgp->varno, vreg, rgp->regno);
rgp++;
}
if(debug['R'] && debug['v'])
dumpit("pass6", firstr);
/*
* pass 7
* peep-hole on basic block
*/
if(!debug['R'] || debug['P']) {
peep();
}
/*
* eliminate nops
* free aux structures
*/
for(p=firstp; p!=P; p=p->link) {
while(p->link != P && p->link->as == ANOP)
p->link = p->link->link;
if(p->to.type == D_BRANCH)
while(p->to.branch != P && p->to.branch->as == ANOP)
p->to.branch = p->to.branch->link;
}
if(r1 != R) {
r1->link = freer;
freer = firstr;
}
if(debug['R']) {
if(ostats.ncvtreg ||
ostats.nspill ||
ostats.nreload ||
ostats.ndelmov ||
ostats.nvar ||
ostats.naddr ||
0)
print("\nstats\n");
if(ostats.ncvtreg)
print(" %4d cvtreg\n", ostats.ncvtreg);
if(ostats.nspill)
print(" %4d spill\n", ostats.nspill);
if(ostats.nreload)
print(" %4d reload\n", ostats.nreload);
if(ostats.ndelmov)
print(" %4d delmov\n", ostats.ndelmov);
if(ostats.nvar)
print(" %4d var\n", ostats.nvar);
if(ostats.naddr)
print(" %4d addr\n", ostats.naddr);
memset(&ostats, 0, sizeof(ostats));
}
}
Bits
mkvar(Reg *r, Adr *a)
{
Var *v;
int i, t, n, et, z, flag;
int64 w;
uint32 regu;
int64 o;
Bits bit;
Node *node;
/*
* mark registers used
*/
t = a->type;
if(t == D_NONE)
goto none;
if(r != R)
r->use1.b[0] |= doregbits(a->index);
switch(t) {
default:
regu = doregbits(t);
if(regu == 0)
goto none;
bit = zbits;
bit.b[0] = regu;
return bit;
case D_ADDR:
a->type = a->index;
bit = mkvar(r, a);
setaddrs(bit);
a->type = t;
ostats.naddr++;
goto none;
case D_EXTERN:
case D_STATIC:
case D_PARAM:
case D_AUTO:
n = t;
break;
}
node = a->node;
if(node == N || node->op != ONAME || node->orig == N)
goto none;
node = node->orig;
if(node->orig != node)
fatal("%D: bad node", a);
if(node->sym == S || node->sym->name[0] == '.')
goto none;
et = a->etype;
o = a->offset;
w = a->width;
if(w < 0)
fatal("bad width %lld for %D", w, a);
flag = 0;
for(i=0; i<nvar; i++) {
v = var+i;
if(v->node == node && v->name == n) {
if(v->offset == o)
if(v->etype == et)
if(v->width == w)
return blsh(i);
// if they overlaps, disable both
if(overlap(v->offset, v->width, o, w)) {
// print("disable overlap %s %d %d %d %d, %E != %E\n", s->name, v->offset, v->width, o, w, v->etype, et);
v->addr = 1;
flag = 1;
}
}
}
switch(et) {
case 0:
case TFUNC:
goto none;
}
if(nvar >= NVAR) {
if(debug['w'] > 1 && node != N)
fatal("variable not optimized: %#N", node);
goto none;
}
i = nvar;
nvar++;
v = var+i;
v->offset = o;
v->name = n;
v->etype = et;
v->width = w;
v->addr = flag; // funny punning
v->node = node;
if(debug['R'])
print("bit=%2d et=%2E w=%lld+%lld %#N %D flag=%d\n", i, et, o, w, node, a, v->addr);
ostats.nvar++;
bit = blsh(i);
if(n == D_EXTERN || n == D_STATIC)
for(z=0; z<BITS; z++)
externs.b[z] |= bit.b[z];
if(n == D_PARAM)
for(z=0; z<BITS; z++)
params.b[z] |= bit.b[z];
return bit;
none:
return zbits;
}
int main(int argc, const char *argv[]) {
/**
* Test 1
*/
struct exp *e1, *e2;
struct explist *l;
l = NULL;
l = cons(mkconstant(5), l);
l = cons(mkconstant(3), l);
l = cons(mkconstant(2), l);
e1 = mkopapp(isplus, l);
l = NULL;
l = cons(mkvar("x"), l);
l = cons(mkvar("x"), l);
l = cons(mkvar("x"), l);
e2 = mkopapp(ismult, l);
e1 = mklet("x", e1, e2);
printf("%d\n", evalexp(e1)); // should print 1000
/**
* Test 2
*/
struct exp *e3, *e4, *e5;
l = NULL;
l = cons(mkconstant(5), l);
l = cons(mkconstant(3), l);
l = cons(mkconstant(2), l);
e1 = mkopapp(isplus, l);
l = NULL;
l = cons(mkvar("y"), l);
l = cons(mkvar("x"), l);
l = cons(mkvar("z"), l);
e2 = mkopapp(ismult, l);
e3 = mklet("x", e1, e2);
e4 = mklet("y", mkconstant(5), e3);
e5 = mklet("z", mkconstant(40), e4);
printf("%d\n", evalexp(e5)); // should print 2000
/**
* Test 3
*/
l = NULL;
l = cons(mkconstant(3), l);
l = cons(mkvar("x"), l);
e2 = mkopapp(isplus, l);
l = NULL;
l = cons(mkconstant(9), l);
l = cons(mkvar("x"), l);
e5 = mkopapp(isplus, l);
e4 = mklet("x", mkconstant(4), e5);
l = NULL;
l = cons(e4, l);
l = cons(mkvar("x"), l);
e3 = mkopapp(isplus, l);
e1 = mklet("x", mklet("x", mkconstant(2), e2), e3);
printf("%d\n", evalexp(e1)); // should print 18
return 0;
}
