michael@0: /*
michael@0:  * jdhuff.h
michael@0:  *
michael@0:  * This file was part of the Independent JPEG Group's software:
michael@0:  * Copyright (C) 1991-1997, Thomas G. Lane.
michael@0:  * Modifications:
michael@0:  * Copyright (C) 2010-2011, D. R. Commander.
michael@0:  * For conditions of distribution and use, see the accompanying README file.
michael@0:  *
michael@0:  * This file contains declarations for Huffman entropy decoding routines
michael@0:  * that are shared between the sequential decoder (jdhuff.c) and the
michael@0:  * progressive decoder (jdphuff.c).  No other modules need to see these.
michael@0:  */
michael@0: 
michael@0: /* Short forms of external names for systems with brain-damaged linkers. */
michael@0: 
michael@0: #ifdef NEED_SHORT_EXTERNAL_NAMES
michael@0: #define jpeg_make_d_derived_tbl	jMkDDerived
michael@0: #define jpeg_fill_bit_buffer	jFilBitBuf
michael@0: #define jpeg_huff_decode	jHufDecode
michael@0: #endif /* NEED_SHORT_EXTERNAL_NAMES */
michael@0: 
michael@0: 
michael@0: /* Derived data constructed for each Huffman table */
michael@0: 
michael@0: #define HUFF_LOOKAHEAD	8	/* # of bits of lookahead */
michael@0: 
michael@0: typedef struct {
michael@0:   /* Basic tables: (element [0] of each array is unused) */
michael@0:   INT32 maxcode[18];		/* largest code of length k (-1 if none) */
michael@0:   /* (maxcode[17] is a sentinel to ensure jpeg_huff_decode terminates) */
michael@0:   INT32 valoffset[18];		/* huffval[] offset for codes of length k */
michael@0:   /* valoffset[k] = huffval[] index of 1st symbol of code length k, less
michael@0:    * the smallest code of length k; so given a code of length k, the
michael@0:    * corresponding symbol is huffval[code + valoffset[k]]
michael@0:    */
michael@0: 
michael@0:   /* Link to public Huffman table (needed only in jpeg_huff_decode) */
michael@0:   JHUFF_TBL *pub;
michael@0: 
michael@0:   /* Lookahead table: indexed by the next HUFF_LOOKAHEAD bits of
michael@0:    * the input data stream.  If the next Huffman code is no more
michael@0:    * than HUFF_LOOKAHEAD bits long, we can obtain its length and
michael@0:    * the corresponding symbol directly from this tables.
michael@0:    *
michael@0:    * The lower 8 bits of each table entry contain the number of
michael@0:    * bits in the corresponding Huffman code, or HUFF_LOOKAHEAD + 1
michael@0:    * if too long.  The next 8 bits of each entry contain the
michael@0:    * symbol.
michael@0:    */
michael@0:   int lookup[1<<HUFF_LOOKAHEAD];
michael@0: } d_derived_tbl;
michael@0: 
michael@0: /* Expand a Huffman table definition into the derived format */
michael@0: EXTERN(void) jpeg_make_d_derived_tbl
michael@0: 	JPP((j_decompress_ptr cinfo, boolean isDC, int tblno,
michael@0: 	     d_derived_tbl ** pdtbl));
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * Fetching the next N bits from the input stream is a time-critical operation
michael@0:  * for the Huffman decoders.  We implement it with a combination of inline
michael@0:  * macros and out-of-line subroutines.  Note that N (the number of bits
michael@0:  * demanded at one time) never exceeds 15 for JPEG use.
michael@0:  *
michael@0:  * We read source bytes into get_buffer and dole out bits as needed.
michael@0:  * If get_buffer already contains enough bits, they are fetched in-line
michael@0:  * by the macros CHECK_BIT_BUFFER and GET_BITS.  When there aren't enough
michael@0:  * bits, jpeg_fill_bit_buffer is called; it will attempt to fill get_buffer
michael@0:  * as full as possible (not just to the number of bits needed; this
michael@0:  * prefetching reduces the overhead cost of calling jpeg_fill_bit_buffer).
michael@0:  * Note that jpeg_fill_bit_buffer may return FALSE to indicate suspension.
michael@0:  * On TRUE return, jpeg_fill_bit_buffer guarantees that get_buffer contains
michael@0:  * at least the requested number of bits --- dummy zeroes are inserted if
michael@0:  * necessary.
michael@0:  */
michael@0: 
michael@0: #if __WORDSIZE == 64 || defined(_WIN64)
michael@0: 
michael@0: typedef size_t bit_buf_type;	/* type of bit-extraction buffer */
michael@0: #define BIT_BUF_SIZE  64		/* size of buffer in bits */
michael@0: 
michael@0: #else
michael@0: 
michael@0: typedef INT32 bit_buf_type;	/* type of bit-extraction buffer */
michael@0: #define BIT_BUF_SIZE  32		/* size of buffer in bits */
michael@0: 
michael@0: #endif
michael@0: 
michael@0: /* If long is > 32 bits on your machine, and shifting/masking longs is
michael@0:  * reasonably fast, making bit_buf_type be long and setting BIT_BUF_SIZE
michael@0:  * appropriately should be a win.  Unfortunately we can't define the size
michael@0:  * with something like  #define BIT_BUF_SIZE (sizeof(bit_buf_type)*8)
michael@0:  * because not all machines measure sizeof in 8-bit bytes.
michael@0:  */
michael@0: 
michael@0: typedef struct {		/* Bitreading state saved across MCUs */
michael@0:   bit_buf_type get_buffer;	/* current bit-extraction buffer */
michael@0:   int bits_left;		/* # of unused bits in it */
michael@0: } bitread_perm_state;
michael@0: 
michael@0: typedef struct {		/* Bitreading working state within an MCU */
michael@0:   /* Current data source location */
michael@0:   /* We need a copy, rather than munging the original, in case of suspension */
michael@0:   const JOCTET * next_input_byte; /* => next byte to read from source */
michael@0:   size_t bytes_in_buffer;	/* # of bytes remaining in source buffer */
michael@0:   /* Bit input buffer --- note these values are kept in register variables,
michael@0:    * not in this struct, inside the inner loops.
michael@0:    */
michael@0:   bit_buf_type get_buffer;	/* current bit-extraction buffer */
michael@0:   int bits_left;		/* # of unused bits in it */
michael@0:   /* Pointer needed by jpeg_fill_bit_buffer. */
michael@0:   j_decompress_ptr cinfo;	/* back link to decompress master record */
michael@0: } bitread_working_state;
michael@0: 
michael@0: /* Macros to declare and load/save bitread local variables. */
michael@0: #define BITREAD_STATE_VARS  \
michael@0: 	register bit_buf_type get_buffer;  \
michael@0: 	register int bits_left;  \
michael@0: 	bitread_working_state br_state
michael@0: 
michael@0: #define BITREAD_LOAD_STATE(cinfop,permstate)  \
michael@0: 	br_state.cinfo = cinfop; \
michael@0: 	br_state.next_input_byte = cinfop->src->next_input_byte; \
michael@0: 	br_state.bytes_in_buffer = cinfop->src->bytes_in_buffer; \
michael@0: 	get_buffer = permstate.get_buffer; \
michael@0: 	bits_left = permstate.bits_left;
michael@0: 
michael@0: #define BITREAD_SAVE_STATE(cinfop,permstate)  \
michael@0: 	cinfop->src->next_input_byte = br_state.next_input_byte; \
michael@0: 	cinfop->src->bytes_in_buffer = br_state.bytes_in_buffer; \
michael@0: 	permstate.get_buffer = get_buffer; \
michael@0: 	permstate.bits_left = bits_left
michael@0: 
michael@0: /*
michael@0:  * These macros provide the in-line portion of bit fetching.
michael@0:  * Use CHECK_BIT_BUFFER to ensure there are N bits in get_buffer
michael@0:  * before using GET_BITS, PEEK_BITS, or DROP_BITS.
michael@0:  * The variables get_buffer and bits_left are assumed to be locals,
michael@0:  * but the state struct might not be (jpeg_huff_decode needs this).
michael@0:  *	CHECK_BIT_BUFFER(state,n,action);
michael@0:  *		Ensure there are N bits in get_buffer; if suspend, take action.
michael@0:  *      val = GET_BITS(n);
michael@0:  *		Fetch next N bits.
michael@0:  *      val = PEEK_BITS(n);
michael@0:  *		Fetch next N bits without removing them from the buffer.
michael@0:  *	DROP_BITS(n);
michael@0:  *		Discard next N bits.
michael@0:  * The value N should be a simple variable, not an expression, because it
michael@0:  * is evaluated multiple times.
michael@0:  */
michael@0: 
michael@0: #define CHECK_BIT_BUFFER(state,nbits,action) \
michael@0: 	{ if (bits_left < (nbits)) {  \
michael@0: 	    if (! jpeg_fill_bit_buffer(&(state),get_buffer,bits_left,nbits))  \
michael@0: 	      { action; }  \
michael@0: 	    get_buffer = (state).get_buffer; bits_left = (state).bits_left; } }
michael@0: 
michael@0: #define GET_BITS(nbits) \
michael@0: 	(((int) (get_buffer >> (bits_left -= (nbits)))) & ((1<<(nbits))-1))
michael@0: 
michael@0: #define PEEK_BITS(nbits) \
michael@0: 	(((int) (get_buffer >> (bits_left -  (nbits)))) & ((1<<(nbits))-1))
michael@0: 
michael@0: #define DROP_BITS(nbits) \
michael@0: 	(bits_left -= (nbits))
michael@0: 
michael@0: /* Load up the bit buffer to a depth of at least nbits */
michael@0: EXTERN(boolean) jpeg_fill_bit_buffer
michael@0: 	JPP((bitread_working_state * state, register bit_buf_type get_buffer,
michael@0: 	     register int bits_left, int nbits));
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * Code for extracting next Huffman-coded symbol from input bit stream.
michael@0:  * Again, this is time-critical and we make the main paths be macros.
michael@0:  *
michael@0:  * We use a lookahead table to process codes of up to HUFF_LOOKAHEAD bits
michael@0:  * without looping.  Usually, more than 95% of the Huffman codes will be 8
michael@0:  * or fewer bits long.  The few overlength codes are handled with a loop,
michael@0:  * which need not be inline code.
michael@0:  *
michael@0:  * Notes about the HUFF_DECODE macro:
michael@0:  * 1. Near the end of the data segment, we may fail to get enough bits
michael@0:  *    for a lookahead.  In that case, we do it the hard way.
michael@0:  * 2. If the lookahead table contains no entry, the next code must be
michael@0:  *    more than HUFF_LOOKAHEAD bits long.
michael@0:  * 3. jpeg_huff_decode returns -1 if forced to suspend.
michael@0:  */
michael@0: 
michael@0: #define HUFF_DECODE(result,state,htbl,failaction,slowlabel) \
michael@0: { register int nb, look; \
michael@0:   if (bits_left < HUFF_LOOKAHEAD) { \
michael@0:     if (! jpeg_fill_bit_buffer(&state,get_buffer,bits_left, 0)) {failaction;} \
michael@0:     get_buffer = state.get_buffer; bits_left = state.bits_left; \
michael@0:     if (bits_left < HUFF_LOOKAHEAD) { \
michael@0:       nb = 1; goto slowlabel; \
michael@0:     } \
michael@0:   } \
michael@0:   look = PEEK_BITS(HUFF_LOOKAHEAD); \
michael@0:   if ((nb = (htbl->lookup[look] >> HUFF_LOOKAHEAD)) <= HUFF_LOOKAHEAD) { \
michael@0:     DROP_BITS(nb); \
michael@0:     result = htbl->lookup[look] & ((1 << HUFF_LOOKAHEAD) - 1); \
michael@0:   } else { \
michael@0: slowlabel: \
michael@0:     if ((result=jpeg_huff_decode(&state,get_buffer,bits_left,htbl,nb)) < 0) \
michael@0: 	{ failaction; } \
michael@0:     get_buffer = state.get_buffer; bits_left = state.bits_left; \
michael@0:   } \
michael@0: }
michael@0: 
michael@0: #define HUFF_DECODE_FAST(s,nb,htbl) \
michael@0:   FILL_BIT_BUFFER_FAST; \
michael@0:   s = PEEK_BITS(HUFF_LOOKAHEAD); \
michael@0:   s = htbl->lookup[s]; \
michael@0:   nb = s >> HUFF_LOOKAHEAD; \
michael@0:   /* Pre-execute the common case of nb <= HUFF_LOOKAHEAD */ \
michael@0:   DROP_BITS(nb); \
michael@0:   s = s & ((1 << HUFF_LOOKAHEAD) - 1); \
michael@0:   if (nb > HUFF_LOOKAHEAD) { \
michael@0:     /* Equivalent of jpeg_huff_decode() */ \
michael@0:     /* Don't use GET_BITS() here because we don't want to modify bits_left */ \
michael@0:     s = (get_buffer >> bits_left) & ((1 << (nb)) - 1); \
michael@0:     while (s > htbl->maxcode[nb]) { \
michael@0:       s <<= 1; \
michael@0:       s |= GET_BITS(1); \
michael@0:       nb++; \
michael@0:     } \
michael@0:     s = htbl->pub->huffval[ (int) (s + htbl->valoffset[nb]) & 0xFF ]; \
michael@0:   }
michael@0: 
michael@0: /* Out-of-line case for Huffman code fetching */
michael@0: EXTERN(int) jpeg_huff_decode
michael@0: 	JPP((bitread_working_state * state, register bit_buf_type get_buffer,
michael@0: 	     register int bits_left, d_derived_tbl * htbl, int min_bits));