path: root/base/gxifast.c

/* Copyright (C) 2001-2023 Artifex Software, Inc.
   All Rights Reserved.

   This software is provided AS-IS with no warranty, either express or
   implied.

   This software is distributed under license and may not be copied,
   modified or distributed except as expressly authorized under the terms
   of the license contained in the file LICENSE in this distribution.

   Refer to licensing information at http://www.artifex.com or contact
   Artifex Software, Inc.,  39 Mesa Street, Suite 108A, San Francisco,
   CA 94129, USA, for further information.
*/


/* Fast monochrome image rendering */
#include "gx.h"
#include "memory_.h"
#include "gpcheck.h"
#include "gsbittab.h"
#include "gserrors.h"
#include "gxfixed.h"
#include "gxarith.h"
#include "gxmatrix.h"
#include "gsccolor.h"
#include "gspaint.h"
#include "gsutil.h"
#include "gxdevice.h"
#include "gxcmap.h"
#include "gxdcolor.h"
#include "gxgstate.h"
#include "gxdevmem.h"
#include "gdevmem.h"		/* for mem_mono_device */
#include "gxcpath.h"
#include "gximage.h"
#include "gzht.h"

#include "valgrind.h"

/* Conditionally include statistics code. */
/* #define COLLECT_STATS_IFAST */

/* ------ Strategy procedure ------ */

/* Check the prototype. */
iclass_proc(gs_image_class_1_simple);

/* Use special fast logic for portrait or landscape black-and-white images. */
static irender_proc(image_render_skip);
static irender_proc(image_render_simple);
static irender_proc(image_render_landscape);

int
gs_image_class_1_simple(gx_image_enum * penum, irender_proc_t *render_fn)
{
    fixed ox = dda_current(penum->dda.pixel0.x);
    fixed oy = dda_current(penum->dda.pixel0.y);

    if (penum->use_rop || penum->spp != 1 || penum->bps != 1)
        return 0;
    switch (penum->posture) {
        case image_portrait:
            {			/* Use fast portrait algorithm. */
                long dev_width =
                    fixed2long_pixround(ox + penum->x_extent.x) -
                    fixed2long_pixround(ox);

                if (dev_width != penum->rect.w) {
                    /*
                     * Add an extra align_bitmap_mod of padding so that
                     * we can align scaled rows with the device.
                     */
                    long line_size =
                        bitmap_raster(any_abs(dev_width)) + align_bitmap_mod;

                    if (penum->adjust != 0 || line_size > max_uint)
                        return 0;
                    /* Must buffer a scan line. */
                    penum->line_width = any_abs(dev_width);
                    penum->line_size = (uint) line_size;
                    penum->line = gs_alloc_bytes(penum->memory,
                                            penum->line_size, "image line");
                    if (penum->line == 0) {
                        return gs_error_VMerror;
                    }
                }
                if_debug2m('b', penum->memory, "[b]render=simple, unpack=copy; rect.w=%d, dev_width=%ld\n",
                           penum->rect.w, dev_width);
                *render_fn = image_render_simple;
                break;
            }
        case image_landscape:
            {			/* Use fast landscape algorithm. */
                long dev_width =
                    fixed2long_pixround(oy + penum->x_extent.y) -
                    fixed2long_pixround(oy);
                long line_size =
                    (dev_width = any_abs(dev_width),
                     bitmap_raster(dev_width) * 8 +
                     ROUND_UP(dev_width, 8) * align_bitmap_mod);

                if ((dev_width != penum->rect.w && penum->adjust != 0) ||
                    line_size > max_uint
                    )
                    return 0;
                /* Must buffer a group of 8N scan lines. */
                penum->line_width = dev_width;
                penum->line_size = (uint) line_size;
                penum->line = gs_alloc_bytes(penum->memory,
                                             penum->line_size, "image line");
                if (penum->line == 0) {
                    return gs_error_VMerror;
                }
#ifdef PACIFY_VALGRIND
                memset(penum->line, 0, penum->line_size); /* For the number of scan lined < 8 */
#endif
                penum->xi_next = penum->line_xy = fixed2int_var_rounded(ox);
                if_debug3m('b', penum->memory,
                           "[b]render=landscape, unpack=copy; rect.w=%d, dev_width=%ld, line_size=%ld\n",
                           penum->rect.w, dev_width, line_size);
                *render_fn = image_render_landscape;
                /* Precompute values needed for rasterizing. */
                penum->dxy =
                    float2fixed(penum->matrix.xy +
                                fixed2float(fixed_epsilon) / 2);
                break;
            }
        default:
            return 0;
    }
    /* Precompute values needed for rasterizing. */
    penum->dxx =
        float2fixed(penum->matrix.xx + fixed2float(fixed_epsilon) / 2);
    /*
     * We don't want to spread the samples, but we have to reset unpack_bps
     * to prevent the buffer pointer from being incremented by 8 bytes per
     * input byte.
     */
    penum->unpack = sample_unpack_copy;
    penum->unpack_bps = 8;
    if (penum->use_mask_color) {
        /*
         * Set the masked color as 'no_color' to make it transparent
         *  according to the mask color range and the decoding.
         */
        penum->masked = true;
        if (penum->mask_color.values[0] == 1) {
            /* if v0 == 1, 1 is transparent since v1 must be == 1 to be a valid range */
            set_nonclient_dev_color(penum->map[0].inverted ? penum->icolor0 : penum->icolor1,
                        gx_no_color_index);
        } else if (penum->mask_color.values[1] == 0) {
            /* if v1 == 0, 0 is transparent since v0 must be == 0 to be a valid range */
            set_nonclient_dev_color(penum->map[0].inverted ? penum->icolor1 : penum->icolor0,
                        gx_no_color_index);
        } else {
            /*
             * The only other possible in-range value is v0 = 0, v1 = 1.
             * The image is completely transparent!
             */
            *render_fn = image_render_skip;
        }
        penum->map[0].decoding = sd_none;
    }
    return 0;
}

/* ------ Rendering procedures ------ */

#define DC_IS_NULL(pdc)\
  (gx_dc_is_pure(pdc) && (pdc)->colors.pure == gx_no_color_index)

/* Skip over a completely transparent image. */
static int
image_render_skip(gx_image_enum * penum, const byte * buffer, int data_x,
                  uint w, int h, gx_device * dev)
{
    return h;
}

/*
 * Scale (and possibly reverse) one scan line of a monobit image.
 * This is used for both portrait and landscape image processing.
 * We pass in an x offset (0 <= line_x < align_bitmap_mod * 8) so that
 * we can align the result with the eventual device X.
 *
 * To be precise, the input to this routine is the w bits starting at
 * bit data_x in buffer.  These w bits expand to abs(x_extent) bits,
 * either inverted (zero = 0xff) or not (zero = 0), starting at bit
 * line_x in line which corresponds to coordinate
 * fixed2int_pixround(xcur + min(x_extent, 0)).  Note that the entire
 * bytes containing the first and last output bits are affected: the
 * other bits in those bytes are set to zero (i.e., the value of the
 * 'zero' argument).
 */
#ifdef COLLECT_STATS_IFAST
struct stats_image_fast_s {
    long
         calls, all0s, all1s, runs, lbit0, byte00, byte01, byte02, byte03,
         byte04, rbit0, lbit1, byte1, rbit1, thin, thin2, nwide, bwide,
         nfill, bfill;
} stats_image_fast;
#  define INCS(stat) ++stats_image_fast.stat
#  define ADDS(stat, n) stats_image_fast.stat += n
#else
#  define INCS(stat) DO_NOTHING
#  define ADDS(stat, n) DO_NOTHING
#endif
static inline void
fill_row(byte *line, int line_x, uint raster, int value)
{
    memset(line + (line_x >> 3), value, raster - (line_x >> 3));
}

static void
image_simple_expand(byte * line, int line_x, uint raster,
                    const byte * buffer, int data_x, uint w,
                    fixed xcur, fixed x_extent, byte zero /* 0 or 0xff */ )
{
    int dbitx = data_x & 7;
    byte sbit = 0x80 >> dbitx;
    byte sbitmask = 0xff >> dbitx;
    uint wx = dbitx + w;
    gx_dda_fixed xl;
    gx_dda_step_fixed dxx4, dxx8, dxx16, dxx24, dxx32;
    register const byte *psrc = buffer + (data_x >> 3);

    /*
     * The following 3 variables define the end of the input data row.
     * We would put them in a struct, except that no compiler that we
     * know of will optimize individual struct members as though they
     * were simple variables (e.g., by putting them in registers).
     *
     * endp points to the byte that contains the bit just beyond the
     * end of the row.  endx gives the bit number of this bit within
     * the byte, with 0 being the *least* significant bit.  endbit is
     * a mask for this bit.
     */
    const byte *endp = psrc + (wx >> 3);
    int endx = ~wx & 7;
    byte endbit = 1 << endx;

    /*
     * The following 3 variables do the same for start of the last run
     * of the input row (think of it as a pointer to just beyond the
     * end of the next-to-last run).
     */
    const byte *stop = endp;
    int stopx;
    byte stopbit = endbit;
    byte data;
    byte one = ~zero;
    fixed xl0;
#ifdef PACIFY_VALGRIND
    byte vbits;
#endif

    if (w == 0)
        return;
    INCS(calls);

    /* Scan backward for the last transition. */
    if (stopbit == 0x80)
        --stop, stopbit = 1;
    else
        stopbit <<= 1;
    /* Now (stop, stopbit) give the last bit of the row. */
#ifdef PACIFY_VALGRIND
    /* Here, we are dealing with a row of bits, rather than bytes.
     * If the width of the bits is not a multiple of 8, we don't
     * fill out the last byte, and valgrind (correctly) tracks the
     * bits in that last byte as being a mix of defined and undefined.
     * When we are scanning through the row bitwise, everything works
     * fine, but our "skip whole bytes" code can confuse valgrind.
     * We know that we won't match the "data == 0xff" for the final
     * byte (i.e. the undefinedness of some of the bits doesn't matter
     * to the correctness of the routine), but valgrind is not smart
     * enough to realise that we know this. Accordingly, we get a false
     * positive "undefined memory read".
     * How do we fix this? Well, one way would be to read in the
     * partial last byte, and explicitly set the undefined bits to
     * be 0.
     *   *stop &= ~(stopbit-1);
     * Unfortunately, stop is a const *, so we can't do that (we could
     * break const, but it is just conceivable that the caller might
     * pass the next string of bits out in a later call, and so we
     * might be corrupting valid data).
     * Instead, we make a call to a valgrind helper. */
    VALGRIND_GET_VBITS(stop,&vbits,1);
    if ((vbits & stopbit)==0) { /* At least our stop bit must be addressable already! */
      byte zero = 0;
      VALGRIND_SET_VBITS(stop,&zero,1);
    }
#endif
    {
        byte stopmask = -stopbit << 1;
        byte last = *stop;

        if (stop == psrc)	/* only 1 input byte */
            stopmask &= sbitmask;
        if (last & stopbit) {
            /* The last bit is a 1: look for a 0-to-1 transition. */
            if (~last & stopmask) {	/* Transition in last byte. */
                last |= stopbit - 1;
            } else {		/* No transition in the last byte. */
                while (stop > psrc && stop[-1] == 0xff)
                    --stop;
                if (stop == psrc ||
                    (stop == psrc + 1 && !(~*psrc & sbitmask))
                    ) {
                    /* The input is all 1s.  Clear the row and exit. */
                    INCS(all1s);
                    fill_row(line, line_x, raster, one);
                    goto end;
                }
                last = *--stop;
            }
            stopx = byte_bit_run_length_0[byte_reverse_bits[last]] - 1;
        } else {
            /* The last bit is a 0: look for a 1-to-0 transition. */
            if (last & stopmask) {	/* Transition in last byte. */
                last &= -stopbit;
            } else {		/* No transition in the last byte. */
                while (stop > psrc && stop[-1] == 0)
                    --stop;
                if (stop == psrc ||
                    (stop == psrc + 1 && !(*psrc & sbitmask))
                    ) {
                    /* The input is all 0s.  Clear the row and exit. */
                    INCS(all0s);
                    fill_row(line, line_x, raster, zero);
                    goto end;
                }
                last = *--stop;
            }
            stopx = byte_bit_run_length_0[byte_reverse_bits[last ^ 0xff]] - 1;
        }
        if (stopx < 0)
            stopx = 7, ++stop;
        stopbit = 1 << stopx;
    }

    /* Pre-clear the row. */
    fill_row(line, line_x, raster, zero);


    /* Extreme negative values of x_extent cause the xl0 calculation
     * to explode. Workaround this here. */
    if (x_extent < min_int + 0x100)
      x_extent += 0x100;

    /* Set up the DDAs. */
    xl0 =
        (x_extent >= 0 ?
         fixed_fraction(fixed_pre_pixround(xcur)) :
         fixed_fraction(fixed_pre_pixround(xcur + x_extent)) - x_extent);
    xl0 += int2fixed(line_x);
    /* We should never get a negative x10 here. If we do, all bets are off. */
    if (xl0 < 0)
        xl0 = 0, x_extent = 0;
    dda_init(xl, xl0, x_extent, w);
    dxx4 = xl.step;
    dda_step_add(dxx4, xl.step);
    /* egcc - 2.91.66 generates incorrect code for
     * dda_step_add(dxx4, dxx4);
     * Using the temp variable.
     */
    dxx8 = dxx4;
    dda_step_add(dxx4, dxx8);
    dxx8 = dxx4;
    dda_step_add(dxx8, dxx4);
    dxx16 = dxx8;
    dda_step_add(dxx16, dxx8);
    dxx24 = dxx16;
    dda_step_add(dxx24, dxx8);
    dxx32 = dxx24;
    dda_step_add(dxx32, dxx8);

    /*
     * Loop invariants:
     *      data = *psrc;
     *      sbit = 1 << n, 0<=n<=7.
     */
    for (data = *psrc;;) {
        int x0, n, bit;
        byte *bp;
        static const byte lmasks[9] = {
            0xff, 0x7f, 0x3f, 0x1f, 0xf, 7, 3, 1, 0
        };
        static const byte rmasks[9] = {
            0, 0x80, 0xc0, 0xe0, 0xf0, 0xf8, 0xfc, 0xfe, 0xff
        };

        INCS(runs);

        /* Scan a run of zeros. */
        data ^= 0xff;		/* invert */
        while (data & sbit) {
            dda_next(xl);
            sbit >>= 1;
            INCS(lbit0);
        }
        if (!sbit) {		/* Scan a run of zero bytes. */
sw:	    if ((data = psrc[1]) != 0) {
                psrc++;
                INCS(byte00);
            } else if ((data = psrc[2]) != 0) {
                dda_state_next(xl.state, dxx8);
                psrc += 2;
                INCS(byte01);
            } else if ((data = psrc[3]) != 0) {
                dda_state_next(xl.state, dxx16);
                psrc += 3;
                INCS(byte02);
            } else if ((data = psrc[4]) != 0) {
                dda_state_next(xl.state, dxx24);
                psrc += 4;
                INCS(byte03);
            } else {
                dda_state_next(xl.state, dxx32);
                psrc += 4;
                INCS(byte04);
                goto sw;
            }
            if (data > 0xf)
                sbit = 0x80;
            else {
                sbit = 0x08;
                dda_state_next(xl.state, dxx4);
            }
            data ^= 0xff;	/* invert */
            while (data & sbit) {
                dda_next(xl);
                sbit >>= 1;
                INCS(rbit0);
            }
        }
        x0 = dda_current_fixed2int(xl);
        if (psrc >= stop && sbit == stopbit) {
            /*
             * We've scanned the last run of 0s.
             * Prepare to fill the final run of 1s.
             * Use int64_t to avoid overflow.
             */
            n = fixed2int((int64_t)xl0 + (int64_t)x_extent) - x0;
        } else {		/* Scan a run of ones. */
            /* We know the current bit is a one. */
            data ^= 0xff;	/* un-invert */
            do {
                dda_next(xl);
                sbit >>= 1;
                INCS(lbit1);
            }
            while (data & sbit);
            if (!sbit) {	/* Scan a run of 0xff bytes. */
                while ((data = *++psrc) == 0xff) {
                    dda_state_next(xl.state, dxx8);
                    INCS(byte1);
                }
                if (data < 0xf0)
                    sbit = 0x80;
                else {
                    sbit = 0x08;
                    dda_state_next(xl.state, dxx4);
                }
                while (data & sbit) {
                    dda_next(xl);
                    sbit >>= 1;
                    INCS(rbit1);
                }
            }
            n = dda_current_fixed2int(xl) - x0;
        }

        /* Fill the run in the scan line. */
        if (n < 0)
            x0 += n, n = -n;
        bp = line + (x0 >> 3);
        bit = x0 & 7;
        if ((n += bit) <= 8) {
            *bp ^= lmasks[bit] - lmasks[n];
            INCS(thin);
        } else if ((n -= 8) <= 8) {
            *bp ^= lmasks[bit];
            bp[1] ^= rmasks[n];
            INCS(thin2);
        } else {
            *bp++ ^= lmasks[bit];
            if (n >= 56) {
                int nb = n >> 3;

                memset(bp, one, nb);
                bp += nb;
                INCS(nwide);
                ADDS(bwide, nb);
            } else {
                ADDS(bfill, n >> 3);
                while ((n -= 8) >= 0)
                    *bp++ = one;
                INCS(nfill);
            }
            *bp ^= rmasks[n & 7];
        }
        if (psrc >= stop && sbit == stopbit)
            break;
    }
 end:
    {
#ifdef PACIFY_VALGRIND
    VALGRIND_SET_VBITS(stop,&vbits,1);
#endif
    }
}

/* Copy one rendered scan line to the device. */
static int
copy_portrait(gx_image_enum * penum, const byte * data, int dx, int raster,
              int x, int y, int w, int h, gx_device * dev)
{
    const gx_device_color *pdc0;
    const gx_device_color *pdc1;
    uint align = ALIGNMENT_MOD(data, align_bitmap_mod);

    /*
     * We know that the lookup table maps 1 bit to 1 bit,
     * so it can only have 2 states: straight-through or invert.
     */
    if (penum->map[0].table.lookup4x1to32[0])
        pdc0 = penum->icolor1, pdc1 = penum->icolor0;
    else
        pdc0 = penum->icolor0, pdc1 = penum->icolor1;
    data -= align;
    dx += align << 3;
    if (gx_dc_is_pure(pdc0) && gx_dc_is_pure(pdc1)) {
        /* Just use copy_mono. */
        dev_proc_copy_mono((*copy_mono)) =
            (h == 1 || (raster & (align_bitmap_mod - 1)) == 0 ?
             dev_proc(dev, copy_mono) : gx_copy_mono_unaligned);
        return (*copy_mono)
            (dev, data, dx, raster, gx_no_bitmap_id,
             x, y, w, h, pdc0->colors.pure, pdc1->colors.pure);
    }
    /*
     * At least one color isn't pure: if the other one is transparent, use
     * the opaque color's fill_masked procedure.  Note that we use a
     * slightly unusual representation for transparent here (per
     * gx_begin_image1): a pure color with pixel value gx_no_color_index.
     */
    {
        const gx_device_color *pdc;
        bool invert;

        if (DC_IS_NULL(pdc1)) {
            pdc = pdc0;
            invert = true;
        } else {
            if (!DC_IS_NULL(pdc0)) {
                int code = gx_device_color_fill_rectangle
                    (pdc0, x, y, w, h, dev, lop_default, NULL);

                if (code < 0)
                    return code;
            }
            pdc = pdc1;
            invert = false;
        }
        return (*pdc->type->fill_masked)
            (pdc, data, dx, raster, gx_no_bitmap_id, x, y, w, h,
             dev, lop_default, invert);

    }
}

/* Rendering procedure for a monobit image with no */
/* skew or rotation and pure colors. */
static int
image_render_simple(gx_image_enum * penum, const byte * buffer, int data_x,
                    uint w, int h, gx_device * dev)
{
    dev_proc_copy_mono((*copy_mono)) = dev_proc(dev, copy_mono);
    const fixed dxx = penum->dxx;
    const byte *line;
    uint line_width, line_size;
    int line_x;
    fixed xcur = dda_current(penum->dda.pixel0.x);
    int ix = fixed2int_pixround(xcur);
    int ixr;
    const int iy = penum->yci, ih = penum->hci;
    gx_device_color * const pdc0 = penum->icolor0;
    gx_device_color * const pdc1 = penum->icolor1;
    int dy;
    int code;

    if (h == 0)
        return 0;
    if ((!DC_IS_NULL(pdc0) &&
         (code = gx_color_load(pdc0, penum->pgs, dev)) < 0) ||
        (!DC_IS_NULL(pdc1) &&
         (code = gx_color_load(pdc1, penum->pgs, dev)) < 0)
        )
        return code;
    if (penum->line == 0) {	/* A direct BitBlt is possible. */
        line = buffer;
        line_size = (w + 7) >> 3;
        line_width = w;
        line_x = 0;
    } else if (copy_mono == mem_mono_copy_mono &&
               dxx > 0 && gx_dc_is_pure(pdc1) && gx_dc_is_pure(pdc0) &&
               /* We know the colors must be (0,1) or (1,0). */
               (pdc0->colors.pure ^ pdc1->colors.pure) == 1 &&
               !penum->clip_image &&
               /*
                * Even if clip_image is false, the clipping rectangle
                * might lie partly outside the device coordinate space
                * if the Margins values are non-zero.
                */
               ix >= 0 &&
               (ixr = fixed2int_pixround(xcur + penum->x_extent.x) - 1) <
                 dev->width &&
               iy >= 0 && iy + ih <= dev->height
        ) {
        /* Do the operation directly into the memory device bitmap. */
        int line_ix;
        int ib_left = ix >> 3, ib_right = ixr >> 3;
        byte *scan_line = scan_line_base((gx_device_memory *) dev, iy);
        byte save_left, save_right, mask;

        line_x = ix & (align_bitmap_mod * 8 - 1);
        line_ix = ix - line_x;
        line_size = (ixr >> 3) + 1 - (line_ix >> 3);
        line_width = ixr + 1 - ix;
        /* We must save and restore any unmodified bits in */
        /* the two edge bytes. */
        save_left = scan_line[ib_left];
        save_right = scan_line[ib_right];
        image_simple_expand(scan_line + (line_ix >> 3), line_x,
                            line_size, buffer, data_x, w, xcur,
                            penum->x_extent.x,
                            (byte)((pdc0->colors.pure == 0) !=
                             (penum->map[0].table.lookup4x1to32[0] == 0) ?
                             0xff : 0));
        if (ix & 7)
            mask = (byte) (0xff00 >> (ix & 7)),
                scan_line[ib_left] =
                (save_left & mask) + (scan_line[ib_left] & ~mask);
        if ((ixr + 1) & 7)
            mask = (byte) (0xff00 >> ((ixr + 1) & 7)),
                scan_line[ib_right] =
                (scan_line[ib_right] & mask) + (save_right & ~mask);
        if (ih <= 1)
            return 1;
        /****** MAY BE UNALIGNED ******/
        line = scan_line + (line_ix >> 3);
        for (dy = 1; dy < ih; dy++) {
            int code = (*copy_mono)
                (dev, line, line_x, line_size, gx_no_bitmap_id,
                 ix, iy + dy, line_width, 1,
                 (gx_color_index)0, (gx_color_index)1);

            if (code < 0)
                return code;
        }
        return 0;
    } else {
        line = penum->line;
        line_size = penum->line_size;
        line_width = penum->line_width;
        line_x = ix & (align_bitmap_mod * 8 - 1);
        image_simple_expand(penum->line, line_x, line_size,
                            buffer, data_x, w, xcur,
                            penum->x_extent.x, 0);
    }

    /* Finally, transfer the scan line to the device. */
    if (dxx < 0)
        ix -= line_width;
    for (dy = 0; dy < ih; dy++) {
        int code = copy_portrait(penum, line, line_x, line_size,
                                 ix, iy + dy, line_width, 1, dev);

        if (code < 0)
            return code;
    }

    return 1;
}

/* Rendering procedure for a 90 degree rotated monobit image */
/* with pure colors.  We buffer and then flip 8 scan lines at a time. */
static int copy_landscape(gx_image_enum *, int, int, bool, gx_device *);
static int
image_render_landscape(gx_image_enum * penum, const byte * buffer, int data_x,
                       uint w, int h, gx_device * dev)
{
    byte *line = penum->line;
    uint raster = bitmap_raster(penum->line_width);
    int ix = penum->xci, iw = penum->wci;
    int xinc, xmod;
    byte *row;
    const byte *orig_row = 0;
    bool y_neg = penum->dxy < 0;

    if (is_fneg(penum->matrix.yx))
        ix += iw, iw = -iw, xinc = -1;
    else
        xinc = 1;
    /*
     * Because of clipping, there may be discontinuous jumps in the values
     * of ix (xci).  If this happens, or if we are at the end of the data or
     * a client has requested flushing, flush the flipping buffer.
     */
    if (ix != penum->xi_next || h == 0) {
        int xi = penum->xi_next;
        int code =
            (xinc > 0 ?
             copy_landscape(penum, penum->line_xy, xi, y_neg, dev) :
             copy_landscape(penum, xi, penum->line_xy, y_neg, dev));

        if (code < 0)
            return code;
        penum->line_xy = penum->xi_next = ix;
        if (h == 0)
            return code;
    }
    for (; iw != 0; iw -= xinc) {
        if (xinc < 0)
            --ix;
        xmod = ix & 7;
        row = line + xmod * raster;
        if (orig_row == 0) {
            image_simple_expand(row, 0, raster,
                                buffer, data_x, w,
                                dda_current(penum->dda.pixel0.y),
                                penum->x_extent.y, 0);
            orig_row = row;
        } else
            memcpy(row, orig_row, raster);
        if (xinc > 0) {
            ++ix;
            if (xmod == 7) {
                int code =
                    copy_landscape(penum, penum->line_xy, ix, y_neg, dev);

                if (code < 0)
                    return code;
                orig_row = 0;
                penum->line_xy = ix;
            }
        } else {
            if (xmod == 0) {
                int code =
                    copy_landscape(penum, ix, penum->line_xy, y_neg, dev);

                if (code < 0)
                    return code;
                orig_row = 0;
                penum->line_xy = ix;
            }
        }
    }
    penum->xi_next = ix;
    return 0;
}

/* Flip and copy one group of scan lines. */
static int
copy_landscape(gx_image_enum * penum, int x0, int x1, bool y_neg,
               gx_device * dev)
{
    byte *line = penum->line;
    uint line_width = penum->line_width;
    uint raster = bitmap_raster(line_width);
    byte *flipped = line + raster * 8;
    int w = x1 - x0;
    int y = fixed2int_pixround(dda_current(penum->dda.pixel0.y));

    if (w == 0 || line_width == 0)
        return 0;
    /* Flip the buffered data from raster x 8 to align_bitmap_mod x */
    /* line_width. */
    if (line_width > 0) {
        int i = (line_width-1)>>3;

#ifdef PACIFY_VALGRIND
        if (line_width & 7) {
            memflip8x8_eol(line + i, raster,
                           flipped + (i << (log2_align_bitmap_mod + 3)),
                           align_bitmap_mod,
                           line_width & 7);
            i--;
        }
#endif

        for (; i >= 0; --i)
            memflip8x8(line + i, raster,
                       flipped + (i << (log2_align_bitmap_mod + 3)),
                       align_bitmap_mod);
    }
    /* Transfer the scan lines to the device. */
    if (w < 0)
        x0 = x1, w = -w;
    if (y_neg)
        y -= line_width;
    return copy_portrait(penum, flipped, x0 & 7, align_bitmap_mod,
                         x0, y, w, line_width, dev);
}