#include "gc.h" Reg*rega(void){ Reg *r; r = freer; if(r == R) { r = alloc(sizeof(*r)); } else freer = r->link; *r = zreg; return r;} intrcmp(const void *a1, const void *a2){ Rgn *p1, *p2; int c1, c2; p1 = (Rgn*)a1; p2 = (Rgn*)a2; c1 = p2->cost; c2 = p1->cost; if(c1 -= c2) return c1; return p2->varno - p1->varno;} voidregopt(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 = RtoB(D_SP) | RtoB(D_AX) | RtoB(D_X0); 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: case AIRETQ: r->p1 = R; r1->s1 = R; } bit = mkvar(r, &p->from); if(bany(&bit)) switch(p->as) { /* * funny */ case ALEAL: case ALEAQ: 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); if(bany(&bit)) switch(p->as) { default: diag(Z, "reg: unknown op: %A", p->as); break; /* * right side read */ case ACMPB: case ACMPL: case ACMPQ: case ACMPW: case ACOMISS: case ACOMISD: case AUCOMISS: case AUCOMISD: for(z=0; z<BITS; z++) r->use2.b[z] |= bit.b[z]; break; /* * right side write */ case ANOP: case AMOVL: case AMOVQ: case AMOVB: case AMOVW: case AMOVBLSX: case AMOVBLZX: case AMOVBQSX: case AMOVBQZX: case AMOVLQSX: case AMOVLQZX: case AMOVWLSX: case AMOVWLZX: case AMOVWQSX: case AMOVWQZX: case AMOVSS: case AMOVSD: case ACVTSD2SL: case ACVTSD2SQ: case ACVTSD2SS: case ACVTSL2SD: case ACVTSL2SS: case ACVTSQ2SD: case ACVTSQ2SS: case ACVTSS2SD: case ACVTSS2SL: case ACVTSS2SQ: case ACVTTSD2SL: case ACVTTSD2SQ: case ACVTTSS2SL: case ACVTTSS2SQ: for(z=0; z<BITS; z++) r->set.b[z] |= bit.b[z]; break; /* * right side read+write */ case AADDB: case AADDL: case AADDQ: case AADDW: case AANDB: case AANDL: case AANDQ: case AANDW: case ASUBB: case ASUBL: case ASUBQ: case ASUBW: case AORB: case AORL: case AORQ: case AORW: case AXORB: case AXORL: case AXORQ: case AXORW: case ASALB: case ASALL: case ASALQ: case ASALW: case ASARB: case ASARL: case ASARQ: case ASARW: case AROLB: case AROLL: case AROLQ: case AROLW: case ARORB: case ARORL: case ARORQ: case ARORW: case ASHLB: case ASHLL: case ASHLQ: case ASHLW: case ASHRB: case ASHRL: case ASHRQ: case ASHRW: case AIMULL: case AIMULQ: case AIMULW: case ANEGL: case ANEGQ: case ANOTL: case ANOTQ: case AADCL: case AADCQ: case ASBBL: case ASBBQ: case AADDSD: case AADDSS: case ACMPSD: case ACMPSS: case ADIVSD: case ADIVSS: case AMAXSD: case AMAXSS: case AMINSD: case AMINSS: case AMULSD: case AMULSS: case ARCPSS: case ARSQRTSS: case ASQRTSD: case ASQRTSS: case ASUBSD: case ASUBSS: case AXORPD: for(z=0; z<BITS; z++) { r->set.b[z] |= bit.b[z]; r->use2.b[z] |= bit.b[z]; } break; /* * funny */ case ACALL: for(z=0; z<BITS; z++) addrs.b[z] |= bit.b[z]; break; } switch(p->as) { case AIMULL: case AIMULQ: case AIMULW: if(p->to.type != D_NONE) break; case AIDIVB: case AIDIVL: case AIDIVQ: case AIDIVW: case AIMULB: case ADIVB: case ADIVL: case ADIVQ: case ADIVW: case AMULB: case AMULL: case AMULQ: case AMULW: case ACWD: case ACDQ: case ACQO: 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 AMOVSQ: case AMOVSW: case ACMPSB: case ACMPSL: case ACMPSQ: case ACMPSW: r->regu |= RtoB(D_SI) | RtoB(D_DI); break; case ASTOSB: case ASTOSL: case ASTOSQ: case ASTOSW: case ASCASB: case ASCASL: case ASCASQ: 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; } } 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)Lines 1–500 of 1356. Load more
#include "gc.h" Reg* rega(void) { Reg *r; r = freer; if(r == R) { r = alloc(sizeof(*r)); } else freer = r->link; *r = zreg; return r; } int rcmp(const void *a1, const void *a2) { Rgn *p1, *p2; int c1, c2; p1 = (Rgn*)a1; p2 = (Rgn*)a2; c1 = p2->cost; c2 = p1->cost; if(c1 -= c2) return c1; return p2->varno - p1->varno; } 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 = RtoB(D_SP) | RtoB(D_AX) | RtoB(D_X0); 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: case AIRETQ: r->p1 = R; r1->s1 = R; } bit = mkvar(r, &p->from); if(bany(&bit)) switch(p->as) { /* * funny */ case ALEAL: case ALEAQ: 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); if(bany(&bit)) switch(p->as) { default: diag(Z, "reg: unknown op: %A", p->as); break; /* * right side read */ case ACMPB: case ACMPL: case ACMPQ: case ACMPW: case ACOMISS: case ACOMISD: case AUCOMISS: case AUCOMISD: for(z=0; z<BITS; z++) r->use2.b[z] |= bit.b[z]; break; /* * right side write */ case ANOP: case AMOVL: case AMOVQ: case AMOVB: case AMOVW: case AMOVBLSX: case AMOVBLZX: case AMOVBQSX: case AMOVBQZX: case AMOVLQSX: case AMOVLQZX: case AMOVWLSX: case AMOVWLZX: case AMOVWQSX: case AMOVWQZX: case AMOVSS: case AMOVSD: case ACVTSD2SL: case ACVTSD2SQ: case ACVTSD2SS: case ACVTSL2SD: case ACVTSL2SS: case ACVTSQ2SD: case ACVTSQ2SS: case ACVTSS2SD: case ACVTSS2SL: case ACVTSS2SQ: case ACVTTSD2SL: case ACVTTSD2SQ: case ACVTTSS2SL: case ACVTTSS2SQ: for(z=0; z<BITS; z++) r->set.b[z] |= bit.b[z]; break; /* * right side read+write */ case AADDB: case AADDL: case AADDQ: case AADDW: case AANDB: case AANDL: case AANDQ: case AANDW: case ASUBB: case ASUBL: case ASUBQ: case ASUBW: case AORB: case AORL: case AORQ: case AORW: case AXORB: case AXORL: case AXORQ: case AXORW: case ASALB: case ASALL: case ASALQ: case ASALW: case ASARB: case ASARL: case ASARQ: case ASARW: case AROLB: case AROLL: case AROLQ: case AROLW: case ARORB: case ARORL: case ARORQ: case ARORW: case ASHLB: case ASHLL: case ASHLQ: case ASHLW: case ASHRB: case ASHRL: case ASHRQ: case ASHRW: case AIMULL: case AIMULQ: case AIMULW: case ANEGL: case ANEGQ: case ANOTL: case ANOTQ: case AADCL: case AADCQ: case ASBBL: case ASBBQ: case AADDSD: case AADDSS: case ACMPSD: case ACMPSS: case ADIVSD: case ADIVSS: case AMAXSD: case AMAXSS: case AMINSD: case AMINSS: case AMULSD: case AMULSS: case ARCPSS: case ARSQRTSS: case ASQRTSD: case ASQRTSS: case ASUBSD: case ASUBSS: case AXORPD: for(z=0; z<BITS; z++) { r->set.b[z] |= bit.b[z]; r->use2.b[z] |= bit.b[z]; } break; /* * funny */ case ACALL: for(z=0; z<BITS; z++) addrs.b[z] |= bit.b[z]; break; } switch(p->as) { case AIMULL: case AIMULQ: case AIMULW: if(p->to.type != D_NONE) break; case AIDIVB: case AIDIVL: case AIDIVQ: case AIDIVW: case AIMULB: case ADIVB: case ADIVL: case ADIVQ: case ADIVW: case AMULB: case AMULL: case AMULQ: case AMULW: case ACWD: case ACDQ: case ACQO: 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 AMOVSQ: case AMOVSW: case ACMPSB: case ACMPSL: case ACMPSQ: case ACMPSW: r->regu |= RtoB(D_SI) | RtoB(D_DI); break; case ASTOSB: case ASTOSL: case ASTOSQ: case ASTOSW: case ASCASB: case ASCASL: case ASCASQ: 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; } } 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)