/*
* substitute s for v in a
* return failure to substitute
*/
int
copysub(Adr *a, Adr *v, Adr *s, int f)
{
int t;
if(copyas(a, v)) {
t = s->type;
if(t >= D_AX && t <= D_R15 || t >= D_X0 && t <= D_X0+15) {
if(f)
a->type = t;
}
return 0;
}
if(regtyp(v)) {
t = v->type;
if(a->type == t+D_INDIR) {
if((s->type == D_BP || s->type == D_R13) && a->index != D_NONE)
return 1; /* can't use BP-base with index */
if(f)
a->type = s->type+D_INDIR;
// return 0;
}
if(a->index == t) {
if(f)
a->index = s->type;
return 0;
}
return 0;
}
return 0;
}
int
smallindir(Addr *a, Addr *reg)
{
return regtyp(reg) &&
a->type == D_INDIR + reg->type &&
a->index == D_NONE &&
0 <= a->offset && a->offset < 4096;
}
/*
* direct reference,
* could be set/use depending on
* semantics
*/
int
copyas(Adr *a, Adr *v)
{
if(regtyp(v))
if(a->type == v->type)
if(a->reg == v->reg)
return 1;
return 0;
}
void
nilopt(Prog *firstp)
{
NilFlow *r;
Prog *p;
Graph *g;
int ncheck, nkill;
g = flowstart(firstp, sizeof(NilFlow));
if(g == nil)
return;
if(debug_checknil > 1 /* || strcmp(curfn->nname->sym->name, "f1") == 0 */)
dumpit("nilopt", g->start, 0);
ncheck = 0;
nkill = 0;
for(r = (NilFlow*)g->start; r != nil; r = (NilFlow*)r->f.link) {
p = r->f.prog;
if(p->as != ACHECKNIL || !regtyp(&p->from))
continue;
ncheck++;
if(stackaddr(&p->from)) {
if(debug_checknil && p->lineno > 1)
warnl(p->lineno, "removed nil check of SP address");
r->kill = 1;
continue;
}
nilwalkfwd(r);
if(r->kill) {
if(debug_checknil && p->lineno > 1)
warnl(p->lineno, "removed nil check before indirect");
continue;
}
nilwalkback(r);
if(r->kill) {
if(debug_checknil && p->lineno > 1)
warnl(p->lineno, "removed repeated nil check");
continue;
}
}
for(r = (NilFlow*)g->start; r != nil; r = (NilFlow*)r->f.link) {
if(r->kill) {
nkill++;
excise(&r->f);
}
}
flowend(g);
if(debug_checknil > 1)
print("%S: removed %d of %d nil checks\n", curfn->nname->sym, nkill, ncheck);
}
/*
* direct reference,
* could be set/use depending on
* semantics
*/
int
copyas(Adr *a, Adr *v)
{
if(a->type != v->type)
return 0;
if(regtyp(v))
return 1;
if(v->type == D_AUTO || v->type == D_PARAM)
if(v->offset == a->offset)
return 1;
return 0;
}
int
copyau1(Prog *p, Adr *v)
{
if(regtyp(v))
if(p->from.type == v->type || p->to.type == v->type)
if(p->reg == v->reg) {
if(a2type(p) != v->type)
print("botch a2type %P\n", p);
return 1;
}
return 0;
}
/*
* either direct or indirect
*/
int
copyau(Adr *a, Adr *v)
{
if(copyas(a, v))
return 1;
if(regtyp(v)) {
if(a->type-D_INDIR == v->type)
return 1;
if(a->index == v->type)
return 1;
}
return 0;
}
static int
regconsttyp(Adr *a)
{
if(regtyp(a))
return 1;
switch(a->type) {
case D_CONST:
case D_FCONST:
case D_SCONST:
case D_ADDR:
return 1;
}
return 0;
}
/*
* direct reference,
* could be set/use depending on
* semantics
*/
static int
copyas(Adr *a, Adr *v)
{
if(D_AL <= a->type && a->type <= D_BL)
fatal("use of byte register");
if(D_AL <= v->type && v->type <= D_BL)
fatal("use of byte register");
if(a->type != v->type)
return 0;
if(regtyp(v))
return 1;
if(v->type == D_AUTO || v->type == D_PARAM)
if(v->offset == a->offset)
return 1;
return 0;
}
/*
* direct reference,
* could be set/use depending on
* semantics
*/
int
copyas(Adr *a, Adr *v)
{
if(regtyp(v)) {
if(a->type == v->type)
if(a->reg == v->reg)
return 1;
} else if(v->type == D_CONST) { /* for constprop */
if(a->type == v->type)
if(a->name == v->name)
if(a->sym == v->sym)
if(a->reg == v->reg)
if(a->offset == v->offset)
return 1;
}
return 0;
}
/*
* compare v to the center
* register in p (p->reg)
* the trick is that this
* register might be D_REG
* D_FREG. there are basically
* two cases,
* ADD r,r,r
* CMP r,r,
*/
static int
copyau1(Prog *p, Adr *v)
{
if(regtyp(v))
if(p->reg == v->reg) {
if(p->to.type != D_NONE) {
if(v->type == p->to.type)
return 1;
return 0;
}
if(p->from.type != D_NONE) {
if(v->type == p->from.type)
return 1;
return 0;
}
print("copyau1: can't tell %P\n", p);
}
return 0;
}
static void
nilwalkfwd(NilFlow *rcheck)
{
NilFlow *r;
Prog *p;
ProgInfo info;
// If the path down from rcheck dereferences the address
// (possibly with a small offset) before writing to memory
// and before any subsequent checks, it's okay to wait for
// that implicit check. Only consider this basic block to
// avoid problems like:
// _ = *x // should panic
// for {} // no writes but infinite loop may be considered visible
for(r = (NilFlow*)uniqs(&rcheck->f); r != nil; r = (NilFlow*)uniqs(&r->f)) {
p = r->f.prog;
proginfo(&info, p);
if((info.flags & LeftRead) && smallindir(&p->from, &rcheck->f.prog->from)) {
rcheck->kill = 1;
return;
}
if((info.flags & (RightRead|RightWrite)) && smallindir(&p->to, &rcheck->f.prog->from)) {
rcheck->kill = 1;
return;
}
// Stop if another nil check happens.
if(p->as == ACHECKNIL)
return;
// Stop if value is lost.
if((info.flags & RightWrite) && sameaddr(&p->to, &rcheck->f.prog->from))
return;
// Stop if memory write.
if((info.flags & RightWrite) && !regtyp(&p->to))
return;
}
}
/*
* either direct or indirect
*/
int
copyau(Adr *a, Adr *v)
{
if(copyas(a, v)) {
if(debug['P'] && debug['v'])
print("\tcopyau: copyas returned 1\n");
return 1;
}
if(regtyp(v)) {
if(a->type-D_INDIR == v->type) {
if(debug['P'] && debug['v'])
print("\tcopyau: found indir use - return 1\n");
return 1;
}
if(a->index == v->type) {
if(debug['P'] && debug['v'])
print("\tcopyau: found index use - return 1\n");
return 1;
}
}
return 0;
}
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(Prog *firstp)
{
Flow *r, *r1;
Graph *g;
Prog *p, *p1;
int t;
g = flowstart(firstp, sizeof(Flow));
if(g == nil)
return;
for(r=g->start, t=0; r!=nil; r=r->link, t++)
r->active = t;
// byte, word arithmetic elimination.
elimshortmov(g);
// constant propagation
// find MOV $con,R followed by
// another MOV $con,R without
// setting R in the interim
for(r=g->start; r!=nil; 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", g->start, 0);
t = 0;
for(r=g->start; r!=nil; 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(g, r)) {
excise(r);
t++;
} else
if(subprop(r) && copyprop(g, r)) {
excise(r);
t++;
}
}
break;
case AMOVBLZX:
case AMOVWLZX:
case AMOVBLSX:
case AMOVWLSX:
if(regtyp(&p->to)) {
r1 = rnops(uniqs(r));
if(r1 != nil) {
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 != nil) {
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=g->start; r!=nil; 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=g->start; r!=nil; 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);
}
}
flowend(g);
}
/*
* the idea is to substitute
* one register for another
* from one MOV to another
* MOV a, R0
* ADD b, R0 / no use of R1
* MOV R0, R1
* would be converted to
* MOV a, R1
* ADD b, R1
* MOV R1, R0
* hopefully, then the former or latter MOV
* will be eliminated by copy propagation.
*/
static int
subprop(Flow *r0)
{
Prog *p;
ProgInfo info;
Adr *v1, *v2;
Flow *r;
int t;
if(debug['P'] && debug['v'])
print("subprop %P\n", r0->prog);
p = r0->prog;
v1 = &p->from;
if(!regtyp(v1)) {
if(debug['P'] && debug['v'])
print("\tnot regtype %D; return 0\n", v1);
return 0;
}
v2 = &p->to;
if(!regtyp(v2)) {
if(debug['P'] && debug['v'])
print("\tnot regtype %D; return 0\n", v2);
return 0;
}
for(r=uniqp(r0); r!=nil; r=uniqp(r)) {
if(debug['P'] && debug['v'])
print("\t? %P\n", r->prog);
if(uniqs(r) == nil) {
if(debug['P'] && debug['v'])
print("\tno unique successor\n");
break;
}
p = r->prog;
proginfo(&info, p);
if(info.flags & Call) {
if(debug['P'] && debug['v'])
print("\tfound %P; return 0\n", p);
return 0;
}
if(info.reguse | info.regset) {
if(debug['P'] && debug['v'])
print("\tfound %P; return 0\n", p);
return 0;
}
if((info.flags & Move) && (info.flags & (SizeL|SizeQ|SizeF|SizeD)) && p->to.type == v1->type)
goto gotit;
if(copyau(&p->from, v2) ||
copyau(&p->to, v2)) {
if(debug['P'] && debug['v'])
print("\tcopyau %D failed\n", v2);
break;
}
if(copysub(&p->from, v1, v2, 0) ||
copysub(&p->to, v1, v2, 0)) {
if(debug['P'] && debug['v'])
print("\tcopysub failed\n");
break;
}
}
if(debug['P'] && debug['v'])
print("\tran off end; return 0\n");
return 0;
gotit:
copysub(&p->to, v1, v2, 1);
if(debug['P']) {
print("gotit: %D->%D\n%P", v1, v2, r->prog);
if(p->from.type == v2->type)
print(" excise");
print("\n");
}
for(r=uniqs(r); r!=r0; r=uniqs(r)) {
p = r->prog;
copysub(&p->from, v1, v2, 1);
copysub(&p->to, v1, v2, 1);
if(debug['P'])
print("%P\n", r->prog);
}
t = v1->type;
v1->type = v2->type;
v2->type = t;
if(debug['P'])
print("%P last\n", r->prog);
return 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;
}
}
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;
}
// movb elimination.
// movb is simulated by the linker
// when a register other than ax, bx, cx, dx
// is used, so rewrite to other instructions
// when possible. a movb into a register
// can smash the entire 32-bit register without
// causing any trouble.
static void
elimshortmov(Reg *r)
{
Prog *p;
USED(r);
for(r=firstr; r!=R; r=r->link) {
p = r->prog;
if(regtyp(&p->to)) {
switch(p->as) {
case AINCB:
case AINCW:
p->as = AINCQ;
break;
case ADECB:
case ADECW:
p->as = ADECQ;
break;
case ANEGB:
case ANEGW:
p->as = ANEGQ;
break;
case ANOTB:
case ANOTW:
p->as = ANOTQ;
break;
}
if(regtyp(&p->from) || p->from.type == D_CONST) {
// move or artihmetic into partial register.
// from another register or constant can be movl.
// we don't switch to 64-bit arithmetic if it can
// change how the carry bit is set (and the carry bit is needed).
switch(p->as) {
case AMOVB:
case AMOVW:
p->as = AMOVQ;
break;
case AADDB:
case AADDW:
if(!needc(p->link))
p->as = AADDQ;
break;
case ASUBB:
case ASUBW:
if(!needc(p->link))
p->as = ASUBQ;
break;
case AMULB:
case AMULW:
p->as = AMULQ;
break;
case AIMULB:
case AIMULW:
p->as = AIMULQ;
break;
case AANDB:
case AANDW:
p->as = AANDQ;
break;
case AORB:
case AORW:
p->as = AORQ;
break;
case AXORB:
case AXORW:
p->as = AXORQ;
break;
case ASHLB:
case ASHLW:
p->as = ASHLQ;
break;
}
} else if(p->from.type >= D_NONE) {
// explicit zero extension, but don't
// do that if source is a byte register
// (only AH can occur and it's forbidden).
switch(p->as) {
case AMOVB:
p->as = AMOVBQZX;
break;
case AMOVW:
p->as = AMOVWQZX;
break;
}
}
}
}
}
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);
}
}
/*
* the idea is to substitute
* one register for another
* from one MOV to another
* MOV a, R0
* ADD b, R0 / no use of R1
* MOV R0, R1
* would be converted to
* MOV a, R1
* ADD b, R1
* MOV R1, R0
* hopefully, then the former or latter MOV
* will be eliminated by copy propagation.
*/
int
subprop(Reg *r0)
{
Prog *p;
Adr *v1, *v2;
Reg *r;
int t;
if(debug['P'] && debug['v'])
print("subprop %P\n", r0->prog);
p = r0->prog;
v1 = &p->from;
if(!regtyp(v1)) {
if(debug['P'] && debug['v'])
print("\tnot regtype %D; return 0\n", v1);
return 0;
}
v2 = &p->to;
if(!regtyp(v2)) {
if(debug['P'] && debug['v'])
print("\tnot regtype %D; return 0\n", v2);
return 0;
}
for(r=uniqp(r0); r!=R; r=uniqp(r)) {
if(debug['P'] && debug['v'])
print("\t? %P\n", r->prog);
if(uniqs(r) == R) {
if(debug['P'] && debug['v'])
print("\tno unique successor\n");
break;
}
p = r->prog;
switch(p->as) {
case ACALL:
if(debug['P'] && debug['v'])
print("\tfound %P; return 0\n", p);
return 0;
case AIMULL:
case AIMULQ:
case AIMULW:
if(p->to.type != D_NONE)
break;
goto giveup;
case ARCLB:
case ARCLL:
case ARCLQ:
case ARCLW:
case ARCRB:
case ARCRL:
case ARCRQ:
case ARCRW:
case AROLB:
case AROLL:
case AROLQ:
case AROLW:
case ARORB:
case ARORL:
case ARORQ:
case ARORW:
case ASALB:
case ASALL:
case ASALQ:
case ASALW:
case ASARB:
case ASARL:
case ASARQ:
case ASARW:
case ASHLB:
case ASHLL:
case ASHLQ:
case ASHLW:
case ASHRB:
case ASHRL:
case ASHRQ:
case ASHRW:
if(p->from.type == D_CONST)
break;
goto giveup;
case ADIVB:
case ADIVL:
case ADIVQ:
case ADIVW:
case AIDIVB:
case AIDIVL:
case AIDIVQ:
case AIDIVW:
case AIMULB:
case AMULB:
case AMULL:
case AMULQ:
case AMULW:
case AREP:
case AREPN:
case ACWD:
case ACDQ:
case ACQO:
case ASTOSB:
case ASTOSL:
case ASTOSQ:
case AMOVSB:
case AMOVSL:
case AMOVSQ:
giveup:
if(debug['P'] && debug['v'])
print("\tfound %P; return 0\n", p);
return 0;
case AMOVL:
case AMOVQ:
case AMOVSS:
case AMOVSD:
if(p->to.type == v1->type)
goto gotit;
break;
}
if(copyau(&p->from, v2) ||
copyau(&p->to, v2)) {
if(debug['P'] && debug['v'])
print("\tcopyau %D failed\n", v2);
break;
}
if(copysub(&p->from, v1, v2, 0) ||
copysub(&p->to, v1, v2, 0)) {
if(debug['P'] && debug['v'])
print("\tcopysub failed\n");
break;
}
}
if(debug['P'] && debug['v'])
print("\tran off end; return 0\n");
return 0;
gotit:
copysub(&p->to, v1, v2, 1);
if(debug['P']) {
print("gotit: %D->%D\n%P", v1, v2, r->prog);
if(p->from.type == v2->type)
print(" excise");
print("\n");
}
for(r=uniqs(r); r!=r0; r=uniqs(r)) {
p = r->prog;
copysub(&p->from, v1, v2, 1);
copysub(&p->to, v1, v2, 1);
if(debug['P'])
print("%P\n", r->prog);
}
t = v1->type;
v1->type = v2->type;
v2->type = t;
if(debug['P'])
print("%P last\n", r->prog);
return 1;
}
/*
* the idea is to substitute
* one register for another
* from one MOV to another
* MOV a, R0
* ADD b, R0 / no use of R1
* MOV R0, R1
* would be converted to
* MOV a, R1
* ADD b, R1
* MOV R1, R0
* hopefully, then the former or latter MOV
* will be eliminated by copy propagation.
*/
static int
subprop(Flow *r0)
{
Prog *p;
Adr *v1, *v2;
Flow *r;
int t;
ProgInfo info;
p = r0->prog;
v1 = &p->from;
if(!regtyp(v1))
return 0;
v2 = &p->to;
if(!regtyp(v2))
return 0;
for(r=uniqp(r0); r!=nil; r=uniqp(r)) {
if(debug['P'] && debug['v'])
print("\t? %P\n", r->prog);
if(uniqs(r) == nil)
break;
p = r->prog;
proginfo(&info, p);
if(info.flags & Call)
return 0;
if(info.reguse | info.regset)
return 0;
if((info.flags & Move) && (info.flags & (SizeL|SizeQ|SizeF|SizeD)) && p->to.type == v1->type)
goto gotit;
if(copyau(&p->from, v2) || copyau(&p->to, v2))
break;
if(copysub(&p->from, v1, v2, 0) || copysub(&p->to, v1, v2, 0))
break;
}
return 0;
gotit:
copysub(&p->to, v1, v2, 1);
if(debug['P']) {
print("gotit: %D->%D\n%P", v1, v2, r->prog);
if(p->from.type == v2->type)
print(" excise");
print("\n");
}
for(r=uniqs(r); r!=r0; r=uniqs(r)) {
p = r->prog;
copysub(&p->from, v1, v2, 1);
copysub(&p->to, v1, v2, 1);
if(debug['P'])
print("%P\n", r->prog);
}
t = v1->type;
v1->type = v2->type;
v2->type = t;
if(debug['P'])
print("%P last\n", r->prog);
return 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;
}
}
// 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;
}
/*
* the idea is to substitute
* one register for another
* from one MOV to another
* MOV a, R0
* ADD b, R0 / no use of R1
* MOV R0, R1
* would be converted to
* MOV a, R1
* ADD b, R1
* MOV R1, R0
* hopefully, then the former or latter MOV
* will be eliminated by copy propagation.
*/
int
subprop(Reg *r0)
{
Prog *p;
Adr *v1, *v2;
Reg *r;
int t;
p = r0->prog;
v1 = &p->from;
if(!regtyp(v1))
return 0;
v2 = &p->to;
if(!regtyp(v2))
return 0;
for(r=uniqp(r0); r!=R; r=uniqp(r)) {
if(uniqs(r) == R)
break;
p = r->prog;
switch(p->as) {
case ACALL:
return 0;
case AIMULL:
case AIMULQ:
case AIMULW:
if(p->to.type != D_NONE)
break;
case ADIVB:
case ADIVL:
case ADIVQ:
case ADIVW:
case AIDIVB:
case AIDIVL:
case AIDIVQ:
case AIDIVW:
case AIMULB:
case AMULB:
case AMULL:
case AMULQ:
case AMULW:
case ARCLB:
case ARCLL:
case ARCLQ:
case ARCLW:
case ARCRB:
case ARCRL:
case ARCRQ:
case ARCRW:
case AROLB:
case AROLL:
case AROLQ:
case AROLW:
case ARORB:
case ARORL:
case ARORQ:
case ARORW:
case ASALB:
case ASALL:
case ASALQ:
case ASALW:
case ASARB:
case ASARL:
case ASARQ:
case ASARW:
case ASHLB:
case ASHLL:
case ASHLQ:
case ASHLW:
case ASHRB:
case ASHRL:
case ASHRQ:
case ASHRW:
case AREP:
case AREPN:
case ACWD:
case ACDQ:
case ACQO:
case ASTOSB:
case ASTOSL:
case ASTOSQ:
case AMOVSB:
case AMOVSL:
case AMOVSQ:
return 0;
case AMOVL:
case AMOVQ:
if(p->to.type == v1->type)
goto gotit;
break;
}
if(copyau(&p->from, v2) ||
copyau(&p->to, v2))
break;
if(copysub(&p->from, v1, v2, 0) ||
copysub(&p->to, v1, v2, 0))
break;
}
return 0;
gotit:
copysub(&p->to, v1, v2, 1);
if(debug['P']) {
print("gotit: %D->%D\n%P", v1, v2, r->prog);
if(p->from.type == v2->type)
print(" excise");
print("\n");
}
for(r=uniqs(r); r!=r0; r=uniqs(r)) {
p = r->prog;
copysub(&p->from, v1, v2, 1);
copysub(&p->to, v1, v2, 1);
if(debug['P'])
print("%P\n", r->prog);
}
t = v1->type;
v1->type = v2->type;
v2->type = t;
if(debug['P'])
print("%P last\n", r->prog);
return 1;
}
int
stackaddr(Addr *a)
{
return regtyp(a) && a->type == D_SP;
}
static void
pushback(Reg *r0)
{
Reg *r, *b;
Prog *p0, *p, t;
b = R;
p0 = r0->prog;
for(r=uniqp(r0); r!=R && uniqs(r)!=R; r=uniqp(r)) {
p = r->prog;
if(p->as != ANOP) {
if(!regconsttyp(&p->from) || !regtyp(&p->to))
break;
if(copyu(p, &p0->to, A) || copyu(p0, &p->to, A))
break;
}
if(p->as == ACALL)
break;
b = r;
}
if(b == R) {
if(debug['v']) {
print("no pushback: %P\n", r0->prog);
if(r)
print("\t%P [%d]\n", r->prog, uniqs(r)!=R);
}
return;
}
if(debug['v']) {
print("pushback\n");
for(r=b;; r=r->link) {
print("\t%P\n", r->prog);
if(r == r0)
break;
}
}
t = *r0->prog;
for(r=uniqp(r0);; r=uniqp(r)) {
p0 = r->link->prog;
p = r->prog;
p0->as = p->as;
p0->lineno = p->lineno;
p0->from = p->from;
p0->to = p->to;
if(r == b)
break;
}
p0 = r->prog;
p0->as = t.as;
p0->lineno = t.lineno;
p0->from = t.from;
p0->to = t.to;
if(debug['v']) {
print("\tafter\n");
for(r=b;; r=r->link) {
print("\t%P\n", r->prog);
if(r == r0)
break;
}
}
}
/*
* the idea is to substitute
* one register for another
* from one MOV to another
* MOV a, R0
* ADD b, R0 / no use of R1
* MOV R0, R1
* would be converted to
* MOV a, R1
* ADD b, R1
* MOV R1, R0
* hopefully, then the former or latter MOV
* will be eliminated by copy propagation.
*/
int
subprop(Reg *r0)
{
Prog *p;
Adr *v1, *v2;
Reg *r;
int t;
p = r0->prog;
v1 = &p->from;
if(!regtyp(v1))
return 0;
v2 = &p->to;
if(!regtyp(v2))
return 0;
for(r=uniqp(r0); r!=R; r=uniqp(r)) {
if(uniqs(r) == R)
break;
p = r->prog;
switch(p->as) {
case AJMPL:
return 0;
case AADD:
case ASUB:
case ASLL:
case ASRL:
case ASRA:
case AOR:
case AAND:
case AXOR:
case AMUL:
case ADIV:
case ADIVL:
case AMOD:
case AMODL:
case AFADDD:
case AFADDF:
case AFSUBD:
case AFSUBF:
case AFMULD:
case AFMULF:
case AFDIVD:
case AFDIVF:
if(p->to.type == v1->type)
if(p->to.reg == v1->reg) {
if(p->reg == NREG)
p->reg = p->to.reg;
goto gotit;
}
break;
case AFMOVF:
case AFMOVD:
case AMOVW:
if(p->to.type == v1->type)
if(p->to.reg == v1->reg)
goto gotit;
break;
}
if(copyau(&p->from, v2) ||
copyau1(p, v2) ||
copyau(&p->to, v2))
break;
if(copysub(&p->from, v1, v2, 0) ||
copysub1(p, v1, v2, 0) ||
copysub(&p->to, v1, v2, 0))
break;
}
return 0;
gotit:
copysub(&p->to, v1, v2, 1);
if(debug['P']) {
print("gotit: %D->%D\n%P", v1, v2, r->prog);
if(p->from.type == v2->type)
print(" excise");
print("\n");
}
for(r=uniqs(r); r!=r0; r=uniqs(r)) {
p = r->prog;
copysub(&p->from, v1, v2, 1);
copysub1(p, v1, v2, 1);
copysub(&p->to, v1, v2, 1);
if(debug['P'])
print("%P\n", r->prog);
}
t = v1->reg;
v1->reg = v2->reg;
v2->reg = t;
if(debug['P'])
print("%P last\n", r->prog);
return 1;
}
// movb elimination.
// movb is simulated by the linker
// when a register other than ax, bx, cx, dx
// is used, so rewrite to other instructions
// when possible. a movb into a register
// can smash the entire 64-bit register without
// causing any trouble.
static void
elimshortmov(Graph *g)
{
Prog *p;
Flow *r;
for(r=g->start; r!=nil; r=r->link) {
p = r->prog;
if(regtyp(&p->to)) {
switch(p->as) {
case AINCB:
case AINCW:
p->as = AINCL;
break;
case ADECB:
case ADECW:
p->as = ADECL;
break;
case ANEGB:
case ANEGW:
p->as = ANEGL;
break;
case ANOTB:
case ANOTW:
p->as = ANOTL;
break;
}
if(regtyp(&p->from) || p->from.type == D_CONST) {
// move or artihmetic into partial register.
// from another register or constant can be movl.
// we don't switch to 32-bit arithmetic if it can
// change how the carry bit is set (and the carry bit is needed).
switch(p->as) {
case AMOVB:
case AMOVW:
p->as = AMOVL;
break;
case AADDB:
case AADDW:
if(!needc(p->link))
p->as = AADDL;
break;
case ASUBB:
case ASUBW:
if(!needc(p->link))
p->as = ASUBL;
break;
case AMULB:
case AMULW:
p->as = AMULL;
break;
case AIMULB:
case AIMULW:
p->as = AIMULL;
break;
case AANDB:
case AANDW:
p->as = AANDL;
break;
case AORB:
case AORW:
p->as = AORL;
break;
case AXORB:
case AXORW:
p->as = AXORL;
break;
case ASHLB:
case ASHLW:
p->as = ASHLL;
break;
}
} else {
// explicit zero extension
switch(p->as) {
case AMOVB:
p->as = AMOVBLZX;
break;
case AMOVW:
p->as = AMOVWLZX;
break;
}
}
}
}
}
void
peep(Prog *firstp)
{
Flow *r, *r1;
Graph *g;
Prog *p, *p1;
int t;
g = flowstart(firstp, sizeof(Flow));
if(g == nil)
return;
for(r=g->start, t=0; r!=nil; r=r->link, t++)
r->active = t;
// byte, word arithmetic elimination.
elimshortmov(g);
// constant propagation
// find MOV $con,R followed by
// another MOV $con,R without
// setting R in the interim
for(r=g->start; r!=nil; r=r->link) {
p = r->prog;
switch(p->as) {
case ALEAL:
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 AMOVSS:
case AMOVSD:
if(regtyp(&p->to))
if(p->from.type == D_CONST)
conprop(r);
break;
}
}
loop1:
if(debug['P'] && debug['v'])
dumpit("loop1", g->start, 0);
t = 0;
for(r=g->start; r!=nil; 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(g, r)) {
excise(r);
t++;
} else
if(subprop(r) && copyprop(g, r)) {
excise(r);
t++;
}
}
break;
case AMOVBLZX:
case AMOVWLZX:
case AMOVBLSX:
case AMOVWLSX:
if(regtyp(&p->to)) {
r1 = rnops(uniqs(r));
if(r1 != nil) {
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=g->start; r!=nil; r=r->link) {
p = r->prog;
if(p->as == AMOVSD)
if(regtyp(&p->from))
if(regtyp(&p->to))
p->as = AMOVAPD;
}
flowend(g);
}
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
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();
}
