diff --git a/vint.c b/vint.c index 5a956e1..86120c4 100644 --- a/vint.c +++ b/vint.c @@ -75,27 +75,26 @@ #define OVERFLOWD(in,n,out,vbmax) if(*in == vbmax) { memcpy(out, in+1, n*(USIZE/8)); return in+1+n*(USIZE/8); } #define OVERFLOWE(in,n,out,op,vbmax) if(op > out + n*(USIZE/8)) { *out = vbmax; memcpy(out+1, in, n*(USIZE/8)); op = out+1+n*(USIZE/8); } -//#define RLE(_ip_,_op_) if(_ip_+1 < e && *_ip_ == *(_ip_+1)) { uint_t *_q = _ip_+1; while(_q+1 < e && *(_q+1) == *_ip_) _q++; unsigned _r = _q - _ip_;\ -// { _ip_+=_r; _r--; *_op_++= _ip_[0]?252:253; vbput32(op, _r); if(_ip_[0]) TEMPLATE2(vbput, USIZE)(op, _ip_[0]); }} -//#define RLE(_ip_,_op_) - unsigned char *TEMPLATE2(vbdec, USIZE)(unsigned char *__restrict in, unsigned n, uint_t *__restrict out) { register uint_t x, *op; OVERFLOWD(in, n, out, VB_MAX); + #define VBE(_i_) TEMPLATE2(_vbget, USIZE)(in, x, op[_i_] = x) for(op = out; op != out+(n&~(UN-1)); op += UN) { VBE(0); VBE(1); VBE(2); VBE(3); __builtin_prefetch(in+8*USIZE, 0); #if UN > 4 VBE(4); VBE(5); VBE(6); VBE(7); #endif } - while(op != out+n) TEMPLATE2(_vbget, USIZE)(in, x, *op++ = x ); + while(op != out+n) + TEMPLATE2(_vbget, USIZE)(in, x, *op++ = x ); return in; } #undef VBE unsigned char *TEMPLATE2(vbenc, USIZE)(uint_t *__restrict in, unsigned n, unsigned char *__restrict out) { - register uint_t x, *ip, *e=in+n; + uint_t x, *ip; unsigned char *op = out; + #define VBD(_i_) x = ip[_i_]; TEMPLATE2(_vbput, USIZE)(op, x, ;); for(ip = in; ip != in+(n&~(UN-1)); ip += UN) { __builtin_prefetch(ip+USIZE*8, 0); VBD(0); VBD(1); VBD(2); VBD(3); @@ -107,6 +106,7 @@ unsigned char *TEMPLATE2(vbenc, USIZE)(uint_t *__restrict in, unsigned n, unsign x = *ip++; TEMPLATE2(_vbput, USIZE)(op, x, ;); } + OVERFLOWE(in,n,out,op,VB_MAX); return op; } @@ -116,7 +116,8 @@ uint_t TEMPLATE2(vbgetx, USIZE)(unsigned char *__restrict in, unsigned idx) { unsigned char *ip; unsigned i; uint_t x; - if(*in == 255) return TEMPLATE2(ctou, USIZE)(in+1+idx*(USIZE/8)); + if(*in == 255) + return TEMPLATE2(ctou, USIZE)(in+1+idx*(USIZE/8)); for(ip = in,i = 0; i <= idx; i++) ip += TEMPLATE2(_vbvlen, USIZE)(*ip); TEMPLATE2(_vbget, USIZE)(in, x, ;); @@ -142,16 +143,17 @@ uint_t TEMPLATE2(vbgetx, USIZE)(unsigned char *__restrict in, unsigned idx) { unsigned char *TEMPLATE2(vbzenc, USIZE)(uint_t *__restrict in, unsigned n, unsigned char *__restrict out, uint_t start) { uint_t *ip,v; unsigned char *op = out; + #define VBZE { v = TEMPLATE2(zigzagenc, USIZE)((*ip)-start); start=*ip++; TEMPLATE2(_vbput, USIZE)(op, v, ;); } for(ip = in; ip != in+(n&~(4-1)); ) { VBZE;VBZE;VBZE;VBZE; } - while(ip < in+n) VBZE; - //OVERFLOWE(in,n,out,op); + while(ip < in+n) VBZE; //OVERFLOWE(in,n,out,op); return op; } #undef VBZE unsigned char *TEMPLATE2(vbzdec, USIZE)(unsigned char *__restrict in, unsigned n, uint_t *__restrict out, uint_t start) { uint_t x,*op; + #define VBZD { TEMPLATE2(_vbget, USIZE)(in, x, ;); *op++ = (start += TEMPLATE2(zigzagdec, USIZE)(x)); } for(op = out; op != out+(n&~(UN-1)); ) { VBZD; VBZD; VBZD; VBZD; #if UN > 4 @@ -159,6 +161,7 @@ unsigned char *TEMPLATE2(vbzdec, USIZE)(unsigned char *__restrict in, unsigned n #endif } while(op != out+n) VBZD; + return in; } #undef VBZD @@ -167,6 +170,7 @@ uint_t TEMPLATE2(vbzgetx, USIZE)(unsigned char *__restrict in, unsigned idx, ui unsigned char *ip; unsigned i; uint_t x; + for(ip = in,i = 0; i <= idx; i++) { TEMPLATE2(_vbget, USIZE)(ip, x, ;); start += x+1; @@ -190,7 +194,7 @@ unsigned TEMPLATE2(vbzgeteq, USIZE)(unsigned char **__restrict in, unsigned n, u unsigned char *TEMPLATE2(VBDENC, USIZE)(uint_t *__restrict in, unsigned n, unsigned char *__restrict out, uint_t start) { unsigned char *op = out; if(!n) return out; - uint_t *ip, b=0,v /*,x=in[0]-start-VDELTA*/; + uint_t *ip, b=0,v; #define VBDE { v = (*ip)-start-VDELTA; start = *ip++; TEMPLATE2(_vbput, USIZE)(op, v, ;); b |= (v /*^ x*/); } for(ip = in; ip != in + (n&~(UN-1)); ) { VBDE;VBDE;VBDE;VBDE; #if UN > 4