void
peep(void)
{
Reg *r, *r1, *r2;
Prog *p, *p1;
int t;
/*
* complete R structure
*/
t = 0;
for(r=firstr; r!=R; r=r1) {
r1 = r->link;
if(r1 == R)
break;
p = r->prog->link;
while(p != r1->prog)
switch(p->as) {
default:
r2 = rega();
r->link = r2;
r2->link = r1;
r2->prog = p;
p->reg = r2;
r2->p1 = r;
r->s1 = r2;
r2->s1 = r1;
r1->p1 = r2;
r = r2;
t++;
case ADATA:
case AGLOBL:
case ANAME:
case ASIGNAME:
case ALOCALS:
case ATYPE:
p = p->link;
}
}
// byte, word arithmetic elimination.
elimshortmov(r);
// constant propagation
// find MOV $con,R followed by
// another MOV $con,R without
// setting R in the interim
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
switch(p->as) {
case ALEAL:
case ALEAQ:
if(regtyp(&p->to))
if(p->from.sym != S)
if(p->from.index == D_NONE || p->from.index == D_CONST)
conprop(r);
break;
case AMOVB:
case AMOVW:
case AMOVL:
case AMOVQ:
case AMOVSS:
case AMOVSD:
if(regtyp(&p->to))
if(p->from.type == D_CONST)
conprop(r);
break;
}
}
loop1:
if(debug['P'] && debug['v'])
dumpit("loop1", firstr);
t = 0;
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
switch(p->as) {
case AMOVL:
case AMOVQ:
case AMOVSS:
case AMOVSD:
if(regtyp(&p->to))
if(regtyp(&p->from)) {
if(copyprop(r)) {
excise(r);
t++;
} else
if(subprop(r) && copyprop(r)) {
excise(r);
t++;
}
}
break;
case AMOVBLZX:
case AMOVWLZX:
case AMOVBLSX:
case AMOVWLSX:
if(regtyp(&p->to)) {
r1 = rnops(uniqs(r));
if(r1 != R) {
p1 = r1->prog;
if(p->as == p1->as && p->to.type == p1->from.type){
p1->as = AMOVL;
t++;
}
}
}
break;
case AMOVBQSX:
case AMOVBQZX:
case AMOVWQSX:
case AMOVWQZX:
case AMOVLQSX:
case AMOVLQZX:
case AMOVQL:
if(regtyp(&p->to)) {
r1 = rnops(uniqs(r));
if(r1 != R) {
p1 = r1->prog;
if(p->as == p1->as && p->to.type == p1->from.type){
p1->as = AMOVQ;
t++;
}
}
}
break;
case AADDL:
case AADDQ:
case AADDW:
if(p->from.type != D_CONST || needc(p->link))
break;
if(p->from.offset == -1){
if(p->as == AADDQ)
p->as = ADECQ;
else
if(p->as == AADDL)
p->as = ADECL;
else
p->as = ADECW;
p->from = zprog.from;
break;
}
if(p->from.offset == 1){
if(p->as == AADDQ)
p->as = AINCQ;
else if(p->as == AADDL)
p->as = AINCL;
else
p->as = AINCW;
p->from = zprog.from;
break;
}
break;
case ASUBL:
case ASUBQ:
case ASUBW:
if(p->from.type != D_CONST || needc(p->link))
break;
if(p->from.offset == -1) {
if(p->as == ASUBQ)
p->as = AINCQ;
else
if(p->as == ASUBL)
p->as = AINCL;
else
p->as = AINCW;
p->from = zprog.from;
break;
}
if(p->from.offset == 1){
if(p->as == ASUBQ)
p->as = ADECQ;
else
if(p->as == ASUBL)
p->as = ADECL;
else
p->as = ADECW;
p->from = zprog.from;
break;
}
break;
}
}
if(t)
goto loop1;
// MOVLQZX removal.
// The MOVLQZX exists to avoid being confused for a
// MOVL that is just copying 32-bit data around during
// copyprop. Now that copyprop is done, remov MOVLQZX R1, R2
// if it is dominated by an earlier ADDL/MOVL/etc into R1 that
// will have already cleared the high bits.
//
// MOVSD removal.
// We never use packed registers, so a MOVSD between registers
// can be replaced by MOVAPD, which moves the pair of float64s
// instead of just the lower one. We only use the lower one, but
// the processor can do better if we do moves using both.
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
if(p->as == AMOVLQZX)
if(regtyp(&p->from))
if(p->from.type == p->to.type)
if(prevl(r, p->from.type))
excise(r);
if(p->as == AMOVSD)
if(regtyp(&p->from))
if(regtyp(&p->to))
p->as = AMOVAPD;
}
// load pipelining
// push any load from memory as early as possible
// to give it time to complete before use.
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
switch(p->as) {
case AMOVB:
case AMOVW:
case AMOVL:
case AMOVQ:
case AMOVLQZX:
if(regtyp(&p->to) && !regconsttyp(&p->from))
pushback(r);
}
}
}
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;
}
}
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));
}
}
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
peep(void)
{
Reg *r, *r1, *r2;
Prog *p, *p1;
int t;
/*
* complete R structure
*/
t = 0;
for(r=firstr; r!=R; r=r1) {
r1 = r->link;
if(r1 == R)
break;
p = r->prog->link;
while(p != r1->prog)
switch(p->as) {
default:
r2 = rega();
r->link = r2;
r2->link = r1;
r2->prog = p;
r2->p1 = r;
r->s1 = r2;
r2->s1 = r1;
r1->p1 = r2;
r = r2;
t++;
case ADATA:
case AGLOBL:
case ANAME:
case ASIGNAME:
p = p->link;
}
}
loop1:
t = 0;
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
if(p->as == ASLL || p->as == ASRL || p->as == ASRA) {
/*
* elide shift into D_SHIFT operand of subsequent instruction
*/
if(shiftprop(r)) {
excise(r);
t++;
}
}
if(p->as == AMOVW || p->as == AMOVF || p->as == AMOVD)
if(regtyp(&p->to)) {
if(p->from.type == D_CONST)
constprop(&p->from, &p->to, r->s1);
else if(regtyp(&p->from))
if(p->from.type == p->to.type) {
if(copyprop(r)) {
excise(r);
t++;
} else
if(subprop(r) && copyprop(r)) {
excise(r);
t++;
}
}
}
}
if(t)
goto loop1;
/*
* look for MOVB x,R; MOVB R,R
*/
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
switch(p->as) {
default:
continue;
case AEOR:
/*
* EOR -1,x,y => MVN x,y
*/
if(p->from.type == D_CONST && p->from.offset == -1) {
p->as = AMVN;
p->from.type = D_REG;
if(p->reg != NREG)
p->from.reg = p->reg;
else
p->from.reg = p->to.reg;
p->reg = NREG;
}
continue;
case AMOVH:
case AMOVHU:
case AMOVB:
case AMOVBU:
if(p->to.type != D_REG)
continue;
break;
}
r1 = r->link;
if(r1 == R)
continue;
p1 = r1->prog;
if(p1->as != p->as)
continue;
if(p1->from.type != D_REG || p1->from.reg != p->to.reg)
continue;
if(p1->to.type != D_REG || p1->to.reg != p->to.reg)
continue;
excise(r1);
}
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
switch(p->as) {
case AMOVW:
case AMOVB:
case AMOVBU:
if(p->from.type == D_OREG && p->from.offset == 0)
xtramodes(r, &p->from);
else if(p->to.type == D_OREG && p->to.offset == 0)
xtramodes(r, &p->to);
else
continue;
break;
case ACMP:
/*
* elide CMP $0,x if calculation of x can set condition codes
*/
if(p->from.type != D_CONST || p->from.offset != 0)
continue;
r2 = r->s1;
if(r2 == R)
continue;
t = r2->prog->as;
switch(t) {
default:
continue;
case ABEQ:
case ABNE:
case ABMI:
case ABPL:
break;
case ABGE:
t = ABPL;
break;
case ABLT:
t = ABMI;
break;
case ABHI:
t = ABNE;
break;
case ABLS:
t = ABEQ;
break;
}
r1 = r;
do
r1 = uniqp(r1);
while (r1 != R && r1->prog->as == ANOP);
if(r1 == R)
continue;
p1 = r1->prog;
if(p1->to.type != D_REG)
continue;
if(p1->to.reg != p->reg)
if(!(p1->as == AMOVW && p1->from.type == D_REG && p1->from.reg == p->reg))
continue;
switch(p1->as) {
default:
continue;
case AMOVW:
if(p1->from.type != D_REG)
continue;
case AAND:
case AEOR:
case AORR:
case ABIC:
case AMVN:
case ASUB:
case ARSB:
case AADD:
case AADC:
case ASBC:
case ARSC:
break;
}
p1->scond |= C_SBIT;
r2->prog->as = t;
excise(r);
continue;
}
}
predicate();
}
void
peep(void)
{
Reg *r, *r1, *r2;
Prog *p, *p1;
int t;
/*
* complete R structure
*/
t = 0;
for(r=firstr; r!=R; r=r1) {
r1 = r->link;
if(r1 == R)
break;
p = r->prog->link;
while(p != r1->prog)
switch(p->as) {
default:
r2 = rega();
r->link = r2;
r2->link = r1;
r2->prog = p;
r2->p1 = r;
r->s1 = r2;
r2->s1 = r1;
r1->p1 = r2;
r = r2;
t++;
case ADATA:
case AGLOBL:
case ANAME:
case ASIGNAME:
p = p->link;
}
}
loop1:
t = 0;
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
if(p->as == AMOVW || p->as == AFMOVF || p->as == AFMOVD)
if(regtyp(&p->to)) {
if(regtyp(&p->from))
if(p->from.type == p->to.type) {
if(copyprop(r)) {
excise(r);
t++;
} else
if(subprop(r) && copyprop(r)) {
excise(r);
t++;
}
}
if(regzer(&p->from))
if(p->to.type == D_REG) {
p->from.type = D_REG;
p->from.reg = 0;
if(copyprop(r)) {
excise(r);
t++;
} else
if(subprop(r) && copyprop(r)) {
excise(r);
t++;
}
}
}
}
if(t)
goto loop1;
/*
* look for MOVB x,R; MOVB R,R
*/
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
switch(p->as) {
default:
continue;
case AMOVH:
case AMOVHU:
case AMOVB:
case AMOVBU:
if(p->to.type != D_REG)
continue;
break;
}
r1 = r->link;
if(r1 == R)
continue;
p1 = r1->prog;
if(p1->as != p->as)
continue;
if(p1->from.type != D_REG || p1->from.reg != p->to.reg)
continue;
if(p1->to.type != D_REG || p1->to.reg != p->to.reg)
continue;
excise(r1);
}
}
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
peep(void)
{
Reg *r, *r1, *r2;
Prog *p, *p1;
int t;
/*
* complete R structure
*/
t = 0;
for(r=firstr; r!=R; r=r1) {
r1 = r->link;
if(r1 == R)
break;
p = r->prog->link;
while(p != r1->prog)
switch(p->as) {
default:
r2 = rega();
r->link = r2;
r2->link = r1;
r2->prog = p;
p->reg = r2;
r2->p1 = r;
r->s1 = r2;
r2->s1 = r1;
r1->p1 = r2;
r = r2;
t++;
case ADATA:
case AGLOBL:
case ANAME:
case ASIGNAME:
p = p->link;
}
}
// constant propagation
// find MOV $con,R followed by
// another MOV $con,R without
// setting R in the interim
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
switch(p->as) {
case ALEAL:
case ALEAQ:
if(regtyp(&p->to))
if(p->from.sym != S)
conprop(r);
break;
case AMOVB:
case AMOVW:
case AMOVL:
case AMOVQ:
case AMOVSS:
case AMOVSD:
if(regtyp(&p->to))
if(p->from.type == D_CONST)
conprop(r);
break;
}
}
loop1:
if(debug['P'] && debug['v'])
dumpit("loop1", firstr);
t = 0;
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
switch(p->as) {
case AMOVL:
case AMOVQ:
case AMOVSS:
case AMOVSD:
if(regtyp(&p->to))
if(regtyp(&p->from)) {
if(copyprop(r)) {
excise(r);
t++;
} else
if(subprop(r) && copyprop(r)) {
excise(r);
t++;
}
}
break;
case AMOVBLZX:
case AMOVWLZX:
case AMOVBLSX:
case AMOVWLSX:
if(regtyp(&p->to)) {
r1 = rnops(uniqs(r));
if(r1 != R) {
p1 = r1->prog;
if(p->as == p1->as && p->to.type == p1->from.type){
p1->as = AMOVL;
t++;
}
}
}
break;
case AMOVBQSX:
case AMOVBQZX:
case AMOVWQSX:
case AMOVWQZX:
case AMOVLQSX:
case AMOVLQZX:
if(regtyp(&p->to)) {
r1 = rnops(uniqs(r));
if(r1 != R) {
p1 = r1->prog;
if(p->as == p1->as && p->to.type == p1->from.type){
p1->as = AMOVQ;
t++;
}
}
}
break;
case AADDL:
case AADDQ:
case AADDW:
if(p->from.type != D_CONST || needc(p->link))
break;
if(p->from.offset == -1){
if(p->as == AADDQ)
p->as = ADECQ;
else
if(p->as == AADDL)
p->as = ADECL;
else
p->as = ADECW;
p->from = zprog.from;
break;
}
if(p->from.offset == 1){
if(p->as == AADDQ)
p->as = AINCQ;
else if(p->as == AADDL)
p->as = AINCL;
else
p->as = AINCW;
p->from = zprog.from;
break;
}
break;
case ASUBL:
case ASUBQ:
case ASUBW:
if(p->from.type != D_CONST || needc(p->link))
break;
if(p->from.offset == -1) {
if(p->as == ASUBQ)
p->as = AINCQ;
else
if(p->as == ASUBL)
p->as = AINCL;
else
p->as = AINCW;
p->from = zprog.from;
break;
}
if(p->from.offset == 1){
if(p->as == ASUBQ)
p->as = ADECQ;
else
if(p->as == ASUBL)
p->as = ADECL;
else
p->as = ADECW;
p->from = zprog.from;
break;
}
break;
}
}
if(t)
goto loop1;
}
void
peep(void)
{
Reg *r, *r1, *r2;
Prog *p, *p1;
int t;
/*
* complete R structure
*/
t = 0;
for(r=firstr; r!=R; r=r1) {
r1 = r->link;
if(r1 == R)
break;
p = r->prog->link;
while(p != r1->prog)
switch(p->as) {
default:
r2 = rega();
r->link = r2;
r2->link = r1;
r2->prog = p;
r2->p1 = r;
r->s1 = r2;
r2->s1 = r1;
r1->p1 = r2;
r = r2;
t++;
case ADATA:
case AGLOBL:
case ANAME:
case ASIGNAME:
p = p->link;
}
}
pc = 0; /* speculating it won't kill */
loop1:
t = 0;
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
switch(p->as) {
case AMOVL:
if(regtyp(&p->to))
if(regtyp(&p->from)) {
if(copyprop(r)) {
excise(r);
t++;
}
if(subprop(r) && copyprop(r)) {
excise(r);
t++;
}
}
break;
case AMOVBLSX:
case AMOVBLZX:
case AMOVWLSX:
case AMOVWLZX:
if(regtyp(&p->to)) {
r1 = uniqs(r);
if(r1 != R) {
p1 = r1->prog;
if(p->as == p1->as && p->to.type == p1->from.type)
p1->as = AMOVL;
}
}
break;
case AADDL:
case AADDW:
if(p->from.type != D_CONST || needc(p->link))
break;
if(p->from.offset == -1) {
if(p->as == AADDL)
p->as = ADECL;
else
p->as = ADECW;
p->from = zprog.from;
}
else if(p->from.offset == 1) {
if(p->as == AADDL)
p->as = AINCL;
else
p->as = AINCW;
p->from = zprog.from;
}
break;
case ASUBL:
case ASUBW:
if(p->from.type != D_CONST || needc(p->link))
break;
if(p->from.offset == -1) {
if(p->as == ASUBL)
p->as = AINCL;
else
p->as = AINCW;
p->from = zprog.from;
}
else if(p->from.offset == 1) {
if(p->as == ASUBL)
p->as = ADECL;
else
p->as = ADECW;
p->from = zprog.from;
}
break;
}
}
if(t)
goto loop1;
}
void
peep(void)
{
Reg *r, *r1, *r2;
Prog *p, *p1;
int t;
/*
* complete R structure
*/
t = 0;
for(r=firstr; r!=R; r=r1) {
r1 = r->link;
if(r1 == R)
break;
p = r->prog->link;
while(p != r1->prog)
switch(p->as) {
default:
r2 = rega();
r->link = r2;
r2->link = r1;
r2->prog = p;
r2->p1 = r;
r->s1 = r2;
r2->s1 = r1;
r1->p1 = r2;
r = r2;
t++;
case ADATA:
case AGLOBL:
case ANAME:
case ASIGNAME:
p = p->link;
}
}
loop1:
t = 0;
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
if(p->as == AMOVW || p->as == AFMOVS || p->as == AFMOVD)
if(regtyp(&p->to)) {
if(regtyp(&p->from))
if(p->from.type == p->to.type) {
if(copyprop(r)) {
excise(r);
t++;
} else
if(subprop(r) && copyprop(r)) {
excise(r);
t++;
}
}
if(regzer(&p->from))
if(p->to.type == D_REG) {
p->from.type = D_REG;
p->from.reg = REGZERO;
if(copyprop(r)) {
excise(r);
t++;
} else
if(subprop(r) && copyprop(r)) {
excise(r);
t++;
}
}
}
}
if(t)
goto loop1;
/*
* look for MOVB x,R; MOVB R,R
*/
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
switch(p->as) {
default:
continue;
case AMOVH:
case AMOVHZ:
case AMOVB:
case AMOVBZ:
if(p->to.type != D_REG)
continue;
break;
}
r1 = r->link;
if(r1 == R)
continue;
p1 = r1->prog;
if(p1->as != p->as)
continue;
if(p1->from.type != D_REG || p1->from.reg != p->to.reg)
continue;
if(p1->to.type != D_REG || p1->to.reg != p->to.reg)
continue;
excise(r1);
}
if(debug['Q'] > 1)
return; /* allow following code improvement to be suppressed */
/*
* look for OP x,y,R; CMP R, $0 -> OPCC x,y,R
* when OP can set condition codes correctly
*/
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
switch(p->as) {
case ACMP:
if(!regzer(&p->to))
continue;
r1 = r->s1;
if(r1 == R)
continue;
switch(r1->prog->as) {
default:
continue;
case ABCL:
case ABC:
/* the conditions can be complex and these are currently little used */
continue;
case ABEQ:
case ABGE:
case ABGT:
case ABLE:
case ABLT:
case ABNE:
case ABVC:
case ABVS:
break;
}
r1 = r;
do
r1 = uniqp(r1);
while (r1 != R && r1->prog->as == ANOP);
if(r1 == R)
continue;
p1 = r1->prog;
if(p1->to.type != D_REG || p1->to.reg != p->from.reg)
continue;
switch(p1->as) {
case ASUB:
case AADD:
case AXOR:
case AOR:
/* irregular instructions */
if(p1->from.type == D_CONST)
continue;
break;
}
switch(p1->as) {
default:
continue;
case AMOVW:
if(p1->from.type != D_REG)
continue;
continue;
case AANDCC:
case AANDNCC:
case AORCC:
case AORNCC:
case AXORCC:
case ASUBCC:
case AADDCC:
t = p1->as;
break;
/* don't deal with floating point instructions for now */
/*
case AFABS: t = AFABSCC; break;
case AFADD: t = AFADDCC; break;
case AFADDS: t = AFADDSCC; break;
case AFCTIW: t = AFCTIWCC; break;
case AFCTIWZ: t = AFCTIWZCC; break;
case AFDIV: t = AFDIVCC; break;
case AFDIVS: t = AFDIVSCC; break;
case AFMADD: t = AFMADDCC; break;
case AFMADDS: t = AFMADDSCC; break;
case AFMOVD: t = AFMOVDCC; break;
case AFMSUB: t = AFMSUBCC; break;
case AFMSUBS: t = AFMSUBSCC; break;
case AFMUL: t = AFMULCC; break;
case AFMULS: t = AFMULSCC; break;
case AFNABS: t = AFNABSCC; break;
case AFNEG: t = AFNEGCC; break;
case AFNMADD: t = AFNMADDCC; break;
case AFNMADDS: t = AFNMADDSCC; break;
case AFNMSUB: t = AFNMSUBCC; break;
case AFNMSUBS: t = AFNMSUBSCC; break;
case AFRSP: t = AFRSPCC; break;
case AFSUB: t = AFSUBCC; break;
case AFSUBS: t = AFSUBSCC; break;
case ACNTLZW: t = ACNTLZWCC; break;
case AMTFSB0: t = AMTFSB0CC; break;
case AMTFSB1: t = AMTFSB1CC; break;
*/
case AADD: t = AADDCC; break;
case AADDV: t = AADDVCC; break;
case AADDC: t = AADDCCC; break;
case AADDCV: t = AADDCVCC; break;
case AADDME: t = AADDMECC; break;
case AADDMEV: t = AADDMEVCC; break;
case AADDE: t = AADDECC; break;
case AADDEV: t = AADDEVCC; break;
case AADDZE: t = AADDZECC; break;
case AADDZEV: t = AADDZEVCC; break;
case AAND: t = AANDCC; break;
case AANDN: t = AANDNCC; break;
case ADIVW: t = ADIVWCC; break;
case ADIVWV: t = ADIVWVCC; break;
case ADIVWU: t = ADIVWUCC; break;
case ADIVWUV: t = ADIVWUVCC; break;
case AEQV: t = AEQVCC; break;
case AEXTSB: t = AEXTSBCC; break;
case AEXTSH: t = AEXTSHCC; break;
case AMULHW: t = AMULHWCC; break;
case AMULHWU: t = AMULHWUCC; break;
case AMULLW: t = AMULLWCC; break;
case AMULLWV: t = AMULLWVCC; break;
case ANAND: t = ANANDCC; break;
case ANEG: t = ANEGCC; break;
case ANEGV: t = ANEGVCC; break;
case ANOR: t = ANORCC; break;
case AOR: t = AORCC; break;
case AORN: t = AORNCC; break;
case AREM: t = AREMCC; break;
case AREMV: t = AREMVCC; break;
case AREMU: t = AREMUCC; break;
case AREMUV: t = AREMUVCC; break;
case ARLWMI: t = ARLWMICC; break;
case ARLWNM: t = ARLWNMCC; break;
case ASLW: t = ASLWCC; break;
case ASRAW: t = ASRAWCC; break;
case ASRW: t = ASRWCC; break;
case ASUB: t = ASUBCC; break;
case ASUBV: t = ASUBVCC; break;
case ASUBC: t = ASUBCCC; break;
case ASUBCV: t = ASUBCVCC; break;
case ASUBME: t = ASUBMECC; break;
case ASUBMEV: t = ASUBMEVCC; break;
case ASUBE: t = ASUBECC; break;
case ASUBEV: t = ASUBEVCC; break;
case ASUBZE: t = ASUBZECC; break;
case ASUBZEV: t = ASUBZEVCC; break;
case AXOR: t = AXORCC; break;
break;
}
if(debug['Q'])
print("cmp %P; %P -> ", p1, p);
p1->as = t;
if(debug['Q'])
print("%P\n", p1);
excise(r);
continue;
}
}
}
void
peep(void)
{
Reg *r, *r1, *r2;
Prog *p, *p1;
int t;
p1 = nil;
USED(p1); // ... in unreachable code...
/*
* complete R structure
*/
for(r=firstr; r!=R; r=r1) {
r1 = r->link;
if(r1 == R)
break;
p = r->prog->link;
while(p != r1->prog)
switch(p->as) {
default:
r2 = rega();
r->link = r2;
r2->link = r1;
r2->prog = p;
r2->p1 = r;
r->s1 = r2;
r2->s1 = r1;
r1->p1 = r2;
r = r2;
case ADATA:
case AGLOBL:
case ANAME:
case ASIGNAME:
p = p->link;
}
}
//dumpit("begin", firstr);
loop1:
t = 0;
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
switch(p->as) {
case ASLL:
case ASRL:
case ASRA:
/*
* elide shift into D_SHIFT operand of subsequent instruction
*/
// if(shiftprop(r)) {
// excise(r);
// t++;
// break;
// }
break;
case AMOVW:
case AMOVF:
case AMOVD:
if(regtyp(&p->from))
if(p->from.type == p->to.type)
if(p->scond == C_SCOND_NONE) {
if(copyprop(r)) {
excise(r);
t++;
break;
}
if(subprop(r) && copyprop(r)) {
excise(r);
t++;
break;
}
}
break;
#ifdef NOTDEF
if(p->scond == C_SCOND_NONE)
if(regtyp(&p->to))
if(isdconst(&p->from)) {
constprop(&p->from, &p->to, r->s1);
}
break;
#endif
}
}
if(t)
goto loop1;
return;
#ifdef NOTDEF
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
switch(p->as) {
// case AEOR:
// /*
// * EOR -1,x,y => MVN x,y
// */
// if(isdconst(&p->from) && p->from.offset == -1) {
// p->as = AMVN;
// p->from.type = D_REG;
// if(p->reg != NREG)
// p->from.reg = p->reg;
// else
// p->from.reg = p->to.reg;
// p->reg = NREG;
// }
// break;
case AMOVH:
case AMOVHU:
case AMOVB:
case AMOVBU:
/*
* look for MOVB x,R; MOVB R,R
*/
if(p->to.type != D_REG)
break;
if(r1 == R)
break;
p1 = r1->prog;
if(p1->as != p->as)
break;
if(p1->from.type != D_REG || p1->from.reg != p->to.reg)
break;
if(p1->to.type != D_REG || p1->to.reg != p->to.reg)
break;
excise(r1);
break;
}
r1 = r->link;
}
// for(r=firstr; r!=R; r=r->link) {
// p = r->prog;
// switch(p->as) {
// case AMOVW:
// case AMOVB:
// case AMOVBU:
// if(p->from.type == D_OREG && p->from.offset == 0)
// xtramodes(r, &p->from);
// else
// if(p->to.type == D_OREG && p->to.offset == 0)
// xtramodes(r, &p->to);
// else
// continue;
// break;
// case ACMP:
// /*
// * elide CMP $0,x if calculation of x can set condition codes
// */
// if(isdconst(&p->from) || p->from.offset != 0)
// continue;
// r2 = r->s1;
// if(r2 == R)
// continue;
// t = r2->prog->as;
// switch(t) {
// default:
// continue;
// case ABEQ:
// case ABNE:
// case ABMI:
// case ABPL:
// break;
// case ABGE:
// t = ABPL;
// break;
// case ABLT:
// t = ABMI;
// break;
// case ABHI:
// t = ABNE;
// break;
// case ABLS:
// t = ABEQ;
// break;
// }
// r1 = r;
// do
// r1 = uniqp(r1);
// while (r1 != R && r1->prog->as == ANOP);
// if(r1 == R)
// continue;
// p1 = r1->prog;
// if(p1->to.type != D_REG)
// continue;
// if(p1->to.reg != p->reg)
// if(!(p1->as == AMOVW && p1->from.type == D_REG && p1->from.reg == p->reg))
// continue;
//
// switch(p1->as) {
// default:
// continue;
// case AMOVW:
// if(p1->from.type != D_REG)
// continue;
// case AAND:
// case AEOR:
// case AORR:
// case ABIC:
// case AMVN:
// case ASUB:
// case ARSB:
// case AADD:
// case AADC:
// case ASBC:
// case ARSC:
// break;
// }
// p1->scond |= C_SBIT;
// r2->prog->as = t;
// excise(r);
// continue;
// }
// }
predicate();
#endif
}
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;
}
}
}
void
peep(void)
{
Reg *r, *r1, *r2;
Prog *p, *p1;
int t;
/*
* complete R structure
*/
t = 0;
for(r=firstr; r!=R; r=r1) {
r1 = r->link;
if(r1 == R)
break;
p = r->prog->link;
while(p != r1->prog)
switch(p->as) {
default:
r2 = rega();
r->link = r2;
r2->link = r1;
r2->prog = p;
p->reg = r2;
r2->p1 = r;
r->s1 = r2;
r2->s1 = r1;
r1->p1 = r2;
r = r2;
t++;
case ADATA:
case AGLOBL:
case ANAME:
case ASIGNAME:
p = p->link;
}
}
// byte, word arithmetic elimination.
elimshortmov(r);
// constant propagation
// find MOV $con,R followed by
// another MOV $con,R without
// setting R in the interim
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
switch(p->as) {
case ALEAL:
if(regtyp(&p->to))
if(p->from.sym != S)
conprop(r);
break;
case AMOVB:
case AMOVW:
case AMOVL:
case AMOVSS:
case AMOVSD:
if(regtyp(&p->to))
if(p->from.type == D_CONST)
conprop(r);
break;
}
}
loop1:
if(debug['P'] && debug['v'])
dumpit("loop1", firstr);
t = 0;
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
switch(p->as) {
case AMOVL:
case AMOVSS:
case AMOVSD:
if(regtyp(&p->to))
if(regtyp(&p->from)) {
if(copyprop(r)) {
excise(r);
t++;
} else
if(subprop(r) && copyprop(r)) {
excise(r);
t++;
}
}
break;
case AMOVBLZX:
case AMOVWLZX:
case AMOVBLSX:
case AMOVWLSX:
if(regtyp(&p->to)) {
r1 = rnops(uniqs(r));
if(r1 != R) {
p1 = r1->prog;
if(p->as == p1->as && p->to.type == p1->from.type){
p1->as = AMOVL;
t++;
}
}
}
break;
case AADDL:
case AADDW:
if(p->from.type != D_CONST || needc(p->link))
break;
if(p->from.offset == -1){
if(p->as == AADDL)
p->as = ADECL;
else
p->as = ADECW;
p->from = zprog.from;
break;
}
if(p->from.offset == 1){
if(p->as == AADDL)
p->as = AINCL;
else
p->as = AINCW;
p->from = zprog.from;
break;
}
break;
case ASUBL:
case ASUBW:
if(p->from.type != D_CONST || needc(p->link))
break;
if(p->from.offset == -1) {
if(p->as == ASUBL)
p->as = AINCL;
else
p->as = AINCW;
p->from = zprog.from;
break;
}
if(p->from.offset == 1){
if(p->as == ASUBL)
p->as = ADECL;
else
p->as = ADECW;
p->from = zprog.from;
break;
}
break;
}
}
if(t)
goto loop1;
// MOVSD removal.
// We never use packed registers, so a MOVSD between registers
// can be replaced by MOVAPD, which moves the pair of float64s
// instead of just the lower one. We only use the lower one, but
// the processor can do better if we do moves using both.
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
if(p->as == AMOVSD)
if(regtyp(&p->from))
if(regtyp(&p->to))
p->as = AMOVAPD;
}
}
