Simple Inner Loop Optimzation

Two simple loop optimizations.

1. Strength reduction of mul[tiplication] by loop induction
variable.

2. Hoisting of (address) base into phi where loop induction
variable is used only as a base (address) offset.

Limited to loops with a single body block, which happily
is always innermost loops. This restriction would not be
very hard to lift - it would require detecting the set of
loop blocks (and ensuring reducibility?)

Limited to loop induction variables with 0 initial value
and increment of 1 (for mul strength reduction). This
limitation is trivial to lift; however all of the
cproc/hare[c]/coremark opportunity is with 0/1 loops for
mul reduction, and 0 initial value for base-offset opt.
This commit is contained in:
Roland Paterson-Jones 2024-12-07 10:22:30 +02:00 committed by Quentin Carbonneaux
parent 348f2eac90
commit 95541ccfb0
4 changed files with 243 additions and 1 deletions

View File

@ -5,7 +5,8 @@ PREFIX = /usr/local
BINDIR = $(PREFIX)/bin BINDIR = $(PREFIX)/bin
COMMOBJ = main.o util.o parse.o abi.o cfg.o mem.o ssa.o alias.o load.o \ COMMOBJ = main.o util.o parse.o abi.o cfg.o mem.o ssa.o alias.o load.o \
copy.o fold.o gvn.o gcm.o simpl.o live.o spill.o rega.o emit.o copy.o fold.o gvn.o gcm.o loopopt.o simpl.o live.o spill.o rega.o \
emit.o
AMD64OBJ = amd64/targ.o amd64/sysv.o amd64/isel.o amd64/emit.o AMD64OBJ = amd64/targ.o amd64/sysv.o amd64/isel.o amd64/emit.o
ARM64OBJ = arm64/targ.o arm64/abi.o arm64/isel.o arm64/emit.o ARM64OBJ = arm64/targ.o arm64/abi.o arm64/isel.o arm64/emit.o
RV64OBJ = rv64/targ.o rv64/abi.o rv64/isel.o rv64/emit.o RV64OBJ = rv64/targ.o rv64/abi.o rv64/isel.o rv64/emit.o

3
all.h
View File

@ -595,6 +595,9 @@ void gvn(Fn *);
int pinned(Ins *); int pinned(Ins *);
void gcm(Fn *); void gcm(Fn *);
/* loopopt.c */
void loopopt(Fn *fn);
/* simpl.c */ /* simpl.c */
void simpl(Fn *); void simpl(Fn *);

236
loopopt.c Normal file
View File

@ -0,0 +1,236 @@
#include "all.h"
static int
forloop(Blk *bh, Blk *bb)
{
if (bh->npred == 2)
if (bh->jmp.type == Jjnz)
if (bb->npred == 1)
if (bb->jmp.type == Jjmp)
if (bb->jmp.s[0].b == bh)
return 1;
return 0;
}
static int
doloop(Blk *bh)
{
if (bh->npred == 2)
if (bh->jmp.type == Jjnz)
if (bh->jmp.s[0].b == bh || bh->jmp.s[1].b == bh)
return 1;
return 0;
}
static int
loopvar(Fn *fn, Blk *bh, Blk *bb, uint np, Ref *prl, Ref *pr0, Ref *pri)
{
Blk *bp;
Phi *p;
Suc *sb, *sp;
Tmp *t;
Ins *i;
assert(bh->npred == 2);
assert(bh->pred[0] == bb || bh->pred[1] == bb);
bp = bh->pred[bh->pred[0] == bb]; /* pre-header */
assert(np < bh->nphi);
p = &bh->phi[np];
if (KBASE(p->cls) != 0)
return 0; /* not integer */
sp = getsuc(bp, bh);
assert(sp->nr == bh->nphi);
*pr0 = sp->r[np];
sb = getsuc(bb, bh);
assert(sb->nr == bh->nphi);
*prl = sb->r[np];
if (rtype(*prl) != RTmp)
return 0; /* loop var must be an incremented tmp */
t = &fn->tmp[prl->val];
if (t->bid != bb->id || t->def == 0)
return 0; /* loop var must be defined by an add ins in the body */
i = t->def;
if (i->op != Oadd || req(i->arg[0], p->to) == req(i->arg[1], p->to))
return 0;
*pri = i->arg[req(i->arg[0], p->to)];
assert(p->cls == i->cls);
if (rtype(*pri) == RTmp) {
t = &fn->tmp[pri->val];
if (t->bid == bh->id || t->bid == bb->id)
return 0;
return 1;
}
else {
assert(rtype(*pri) == RCon);
return 1;
}
return 0;
}
static void
mulredux(Fn *fn, Blk *bh, Blk *bb, Blk *bp, uint np, Ref rl, Ref r0, Ref ri, Ins **pvins, uint *pnins)
{
Ins *i;
Tmp *t;
Ref r1, rlop;
Phi *p;
Suc *sb, *sp;
/* dead simple case for now... majority case in practice */
if (!req(r0, con01[0]) || !req(ri, con01[1]))
return;
assert(!req(rl, R)); /* rl unused */
assert(np < bh->nphi);
p = &bh->phi[np];
for (i = bb->ins; i < &bb->ins[bb->nins]; i++) {
/* TODO i->cls != p->cls is not necessary? */
if (i->op != Omul || i->cls != p->cls
|| req(i->arg[0], p->to) == req(i->arg[1], p->to))
continue;
r1 = i->arg[req(i->arg[0], p->to)];
if (rtype(r1) == RTmp) {
t = &fn->tmp[r1.val];
if (t->bid == bh->id || t->bid == bb->id)
continue;
} else
assert(rtype(r1) == RCon);
assert(KBASE(i->cls) == 0);
rlop = newtmp("lop", i->cls, fn);
vgrow(&bh->phi, ++bh->nphi);
bh->phi[bh->nphi-1] = (Phi){.to = i->to, .cls = i->cls};
sp = getsuc(bp, bh);
vgrow(&sp->r, ++sp->nr);
sp->r[sp->nr-1] = con01[0];
assert(sp->nr == bh->nphi);
sb = getsuc(bb, bh);
vgrow(&sb->r, ++sb->nr);
sb->r[sp->nr-1] = rlop;
assert(sb->nr == bh->nphi);
addins(pvins, pnins, &(Ins){.to = rlop, .op = Oadd, .cls = i->cls,
.arg = {i->to, r1}});
*i = (Ins){.op = Onop};
}
}
static void
addbase(Fn *fn, Blk *bh, Blk *bb, Blk *bp, uint np, Ref rl, Ref r0, Ref ri, Ins **pvins, uint *pnins)
{
Ins *i, *ii, *ib;
Tmp *t, *ti, *tb;
Use *u;
Ref rb;
Phi *p;
Suc *sp;
/* dead simple case for now... majority case in practice */
if (!req(r0, con01[0]))
return;
assert(!req(ri, R)); /* unused */
assert(pvins && pnins); /* unused */
assert(np < bh->nphi);
p = &bh->phi[np];
assert(rtype(p->to) == RTmp);
t = &fn->tmp[p->to.val];
if (t->nuse != 2)
return;
i = ib = 0;
for (u = t->use; u < &t->use[t->nuse]; u++) {
if (u->type != UIns)
return;
i = u->u.ins;
if (i->op != Oadd || i->cls != p->cls || u->bid != bb->id)
return;
assert(req(i->arg[0], p->to) || req(i->arg[1], p->to));
if (req(i->to, rl)) {
assert(rtype(i->to) == RTmp);
ti = &fn->tmp[i->to.val];
if (ti->nuse != 1)
return;
assert(ti->use[0].type == UPhi);
assert(ti->use[0].bid == bh->id);
assert(ti->use[0].u.p.np == np);
assert(ti->use[0].u.p.pbid == bb->id);
ii = i;
continue;
}
rb = i->arg[req(i->arg[0], p->to)];
if (req(rb, p->to))
return;
if (rtype(rb) == RTmp) {
tb = &fn->tmp[rb.val];
if (tb->bid == bh->id || tb->bid == bb->id)
return;
} else
assert(rtype(rb) == RCon);
ib = i;
}
assert(ii);
assert(ib);
sp = getsuc(bp, bh);
assert(sp->nr == bh->nphi);
assert(req(sp->r[np], con01[0]));
sp->r[np] = rb;
assert(req(ii->arg[0], p->to) != req(ii->arg[1], p->to));
ii->arg[!req(ii->arg[0], p->to)] = ib->to;
p->to = ib->to;
*ib = (Ins){.op = Onop};
}
typedef void optfn_t(Fn *, Blk *, Blk *, Blk *, uint, Ref, Ref, Ref, Ins **, uint *);
static void
loopvaropt(Fn *fn, optfn_t *optfn)
{
uint bid;
Blk *bh, *bb, *bp; /* header, body, pre-header */
Phi *p;
Ins *vins;
uint nins, np;
Ref rl, r0, ri; /* loop var, loop var init, loop var inc */
nins = 0;
vins = vnew(nins, sizeof vins[0], PFn); /* TODO use insb */
for (bid = 0; bid < fn->nblk; bid++) {
bh = fn->rpo[bid];
if (forloop(bh, bh->jmp.s[0].b))
bb = bh->jmp.s[0].b;
else if (forloop(bh, bh->jmp.s[1].b))
bb = bh->jmp.s[1].b;
else if (doloop(bh))
bb = bh; /* header and body */
else
continue;
bp = bh->pred[bh->pred[0] == bh];
assert(bh->loop > 1);
assert(bh->loop == bb->loop);
nins = 0;
for (np = 0; np < bh->nphi; np++)
if (loopvar(fn, bh, bb, np, &rl, &r0, &ri)) {
p = &bh->phi[np];
assert(KBASE(p->cls) == 0);
optfn(fn, bh, bb, bp, np, rl, r0, ri, &vins, &nins);
}
addnins(&bb->ins, &bb->nins, vins, nins);
}
}
void
loopopt(Fn *fn)
{
/* encourage simple loops */
ifelim(fn);
fillcfg(fn);
blkmerge(fn);
fillcfg(fn);
filluse(fn);
fillloop(fn);
loopvaropt(fn, mulredux);
filluse(fn);
loopvaropt(fn, addbase);
}

2
main.c
View File

@ -81,6 +81,8 @@ func(Fn *fn)
gcm(fn); gcm(fn);
filluse(fn); filluse(fn);
ssacheck(fn); ssacheck(fn);
loopopt(fn);
filluse(fn);
T.abi1(fn); T.abi1(fn);
simpl(fn); simpl(fn);
fillcfg(fn); fillcfg(fn);