+ /* Don't expect this to ever happen: */
+ if (row_width == 0)
+ png_error(png_ptr, "internal row width error");
+
+ /* Preserve the last byte in cases where only part of it will be overwritten,
+ * the multiply below may overflow, we don't care because ANSI-C guarantees
+ * we get the low bits.
+ */
+ end_mask = (pixel_depth * row_width) & 7;
+ if (end_mask != 0)
+ {
+ /* end_ptr == NULL is a flag to say do nothing */
+ end_ptr = dp + PNG_ROWBYTES(pixel_depth, row_width) - 1;
+ end_byte = *end_ptr;
+# ifdef PNG_READ_PACKSWAP_SUPPORTED
+ if (png_ptr->transformations & PNG_PACKSWAP) /* little-endian byte */
+ end_mask = 0xff << end_mask;
+
+ else /* big-endian byte */
+# endif
+ end_mask = 0xff >> end_mask;
+ /* end_mask is now the bits to *keep* from the destination row */
+ }
+
+ /* For non-interlaced images this reduces to a png_memcpy(). A png_memcpy()
+ * will also happen if interlacing isn't supported or if the application
+ * does not call png_set_interlace_handling(). In the latter cases the
+ * caller just gets a sequence of the unexpanded rows from each interlace
+ * pass.
+ */
+#ifdef PNG_READ_INTERLACING_SUPPORTED
+ if (png_ptr->interlaced && (png_ptr->transformations & PNG_INTERLACE) &&
+ pass < 6 && (display == 0 ||
+ /* The following copies everything for 'display' on passes 0, 2 and 4. */
+ (display == 1 && (pass & 1) != 0)))
+ {
+ /* Narrow images may have no bits in a pass; the caller should handle
+ * this, but this test is cheap:
+ */
+ if (row_width <= PNG_PASS_START_COL(pass))
+ return;
+
+ if (pixel_depth < 8)
+ {
+ /* For pixel depths up to 4-bpp the 8-pixel mask can be expanded to fit
+ * into 32 bits, then a single loop over the bytes using the four byte
+ * values in the 32-bit mask can be used. For the 'display' option the
+ * expanded mask may also not require any masking within a byte. To
+ * make this work the PACKSWAP option must be taken into account - it
+ * simply requires the pixels to be reversed in each byte.
+ *
+ * The 'regular' case requires a mask for each of the first 6 passes,
+ * the 'display' case does a copy for the even passes in the range
+ * 0..6. This has already been handled in the tst above.
+ *
+ * The masks are arranged as four bytes with the first byte to use in
+ * the lowest bits (little-endian) regardless of the order (PACKSWAP or
+ * not) of the pixels in each byte.
+ *
+ * NOTE: the whole of this logic depends on the caller of this function
+ * only calling it on rows appropriate to the pass. This function only
+ * understands the 'x' logic, the 'y' logic is handled by the caller.
+ *
+ * The following defines allow generation of compile time constant bit
+ * masks for each pixel depth and each possibility of swapped or not
+ * swapped bytes. Pass 'p' is in the range 0..6; 'x', a pixel index,
+ * is in the range 0..7; and the result is 1 if the pixel is to be
+ * copied in the pass, 0 if not. 'S' is for the sparkle method, 'B'
+ * for the block method.
+ *
+ * With some compilers a compile time expression of the general form:
+ *
+ * (shift >= 32) ? (a >> (shift-32)) : (b >> shift)
+ *
+ * Produces warnings with values of 'shift' in the range 33 to 63
+ * because the right hand side of the ?: expression is evaluated by
+ * the compiler even though it isn't used. Microsoft Visual C (various
+ * versions) and the Intel C compiler are known to do this. To avoid
+ * this the following macros are used in 1.5.6. This is a temporary
+ * solution to avoid destablizing the code during the release process.
+ */
+# if PNG_USE_COMPILE_TIME_MASKS
+# define PNG_LSR(x,s) ((x)>>((s) & 0x1f))
+# define PNG_LSL(x,s) ((x)<<((s) & 0x1f))
+# else
+# define PNG_LSR(x,s) ((x)>>(s))
+# define PNG_LSL(x,s) ((x)<<(s))
+# endif
+# define S_COPY(p,x) (((p)<4 ? PNG_LSR(0x80088822,(3-(p))*8+(7-(x))) :\
+ PNG_LSR(0xaa55ff00,(7-(p))*8+(7-(x)))) & 1)
+# define B_COPY(p,x) (((p)<4 ? PNG_LSR(0xff0fff33,(3-(p))*8+(7-(x))) :\
+ PNG_LSR(0xff55ff00,(7-(p))*8+(7-(x)))) & 1)
+
+ /* Return a mask for pass 'p' pixel 'x' at depth 'd'. The mask is
+ * little endian - the first pixel is at bit 0 - however the extra
+ * parameter 's' can be set to cause the mask position to be swapped
+ * within each byte, to match the PNG format. This is done by XOR of
+ * the shift with 7, 6 or 4 for bit depths 1, 2 and 4.
+ */
+# define PIXEL_MASK(p,x,d,s) \
+ (PNG_LSL(((PNG_LSL(1U,(d)))-1),(((x)*(d))^((s)?8-(d):0))))
+
+ /* Hence generate the appropriate 'block' or 'sparkle' pixel copy mask.
+ */
+# define S_MASKx(p,x,d,s) (S_COPY(p,x)?PIXEL_MASK(p,x,d,s):0)
+# define B_MASKx(p,x,d,s) (B_COPY(p,x)?PIXEL_MASK(p,x,d,s):0)
+
+ /* Combine 8 of these to get the full mask. For the 1-bpp and 2-bpp
+ * cases the result needs replicating, for the 4-bpp case the above
+ * generates a full 32 bits.
+ */
+# define MASK_EXPAND(m,d) ((m)*((d)==1?0x01010101:((d)==2?0x00010001:1)))
+
+# define S_MASK(p,d,s) MASK_EXPAND(S_MASKx(p,0,d,s) + S_MASKx(p,1,d,s) +\
+ S_MASKx(p,2,d,s) + S_MASKx(p,3,d,s) + S_MASKx(p,4,d,s) +\
+ S_MASKx(p,5,d,s) + S_MASKx(p,6,d,s) + S_MASKx(p,7,d,s), d)
+
+# define B_MASK(p,d,s) MASK_EXPAND(B_MASKx(p,0,d,s) + B_MASKx(p,1,d,s) +\
+ B_MASKx(p,2,d,s) + B_MASKx(p,3,d,s) + B_MASKx(p,4,d,s) +\
+ B_MASKx(p,5,d,s) + B_MASKx(p,6,d,s) + B_MASKx(p,7,d,s), d)
+
+#if PNG_USE_COMPILE_TIME_MASKS
+ /* Utility macros to construct all the masks for a depth/swap
+ * combination. The 's' parameter says whether the format is PNG
+ * (big endian bytes) or not. Only the three odd numbered passes are
+ * required for the display/block algorithm.
+ */
+# define S_MASKS(d,s) { S_MASK(0,d,s), S_MASK(1,d,s), S_MASK(2,d,s),\
+ S_MASK(3,d,s), S_MASK(4,d,s), S_MASK(5,d,s) }
+
+# define B_MASKS(d,s) { B_MASK(1,d,s), S_MASK(3,d,s), S_MASK(5,d,s) }
+
+# define DEPTH_INDEX(d) ((d)==1?0:((d)==2?1:2))
+
+ /* Hence the pre-compiled masks indexed by PACKSWAP (or not), depth and
+ * then pass:
+ */
+ static PNG_CONST png_uint_32 row_mask[2/*PACKSWAP*/][3/*depth*/][6] = {
+ /* Little-endian byte masks for PACKSWAP */
+ { S_MASKS(1,0), S_MASKS(2,0), S_MASKS(4,0) },
+ /* Normal (big-endian byte) masks - PNG format */
+ { S_MASKS(1,1), S_MASKS(2,1), S_MASKS(4,1) }
+ };
+
+ /* display_mask has only three entries for the odd passes, so index by
+ * pass>>1.
+ */
+ static PNG_CONST png_uint_32 display_mask[2][3][3] = {
+ /* Little-endian byte masks for PACKSWAP */
+ { B_MASKS(1,0), B_MASKS(2,0), B_MASKS(4,0) },
+ /* Normal (big-endian byte) masks - PNG format */
+ { B_MASKS(1,1), B_MASKS(2,1), B_MASKS(4,1) }
+ };
+
+# define MASK(pass,depth,display,png)\
+ ((display)?display_mask[png][DEPTH_INDEX(depth)][pass>>1]:\
+ row_mask[png][DEPTH_INDEX(depth)][pass])
+
+#else /* !PNG_USE_COMPILE_TIME_MASKS */
+ /* This is the runtime alternative: it seems unlikely that this will
+ * ever be either smaller or faster than the compile time approach.
+ */
+# define MASK(pass,depth,display,png)\
+ ((display)?B_MASK(pass,depth,png):S_MASK(pass,depth,png))
+#endif /* !PNG_USE_COMPILE_TIME_MASKS */
+
+ /* Use the appropriate mask to copy the required bits. In some cases
+ * the byte mask will be 0 or 0xff, optimize these cases. row_width is
+ * the number of pixels, but the code copies bytes, so it is necessary
+ * to special case the end.
+ */
+ png_uint_32 pixels_per_byte = 8 / pixel_depth;
+ png_uint_32 mask;
+
+# ifdef PNG_READ_PACKSWAP_SUPPORTED