//byte* p = (byte*)outptr - dist;
      //byte* q = &p[len];
      //for( p=p; p<q; p++ ) put0( *p );

      if( __builtin_expect(dist>=len,1) ) {
        byte* restrict u = outptr - dist;
        byte* restrict v = outptr;

        if( __builtin_expect(((u-((byte*)0))&7)==0,0) ) {
          __assume_aligned(u,8);
          if( __builtin_expect(((v-((byte*)0))&7)==0,0) ) {
            __assume_aligned(v,8);
            for( i=0; __builtin_expect(i<len,0); i+=8 ) (qword&)v[i]=(qword&)u[i];
          } else for( i=0; __builtin_expect(i<len,0); i+=8 ) (qword&)v[i]=(qword&)u[i];
        } else
        if( __builtin_expect(((v-((byte*)0))&7)==0,0) ) {
          __assume_aligned(v,8);
          for( i=0; __builtin_expect(i<len,0); i+=8 ) (qword&)v[i]=(qword&)u[i];
        } else for( i=0; __builtin_expect(i<len,0); i+=8 ) (qword&)v[i]=(qword&)u[i];

        for( i=i; __builtin_expect(i<len,0); i+=1 ) v[i]=u[i];

      } else {
        byte* u = outptr - dist;
        byte* v = outptr;
        for( i=i; __builtin_expect(i<len,1); i+=1 ) v[i]=u[i];
      }
      outptr+=len;