Bug 689818

The exception occurs in gdevm1.c at line 493:

run -sOutputFile=test.ppm -sDEVICE=pbmraw -dMaxBitmap=10000 ./imagemask-with-pattern.ps
...
Program received signal EXC_ARITHMETIC, Arithmetic exception.
0x0046c6ae in mem_mono_strip_tile_rectangle (dev=0x176e6c8, tiles=0xbfffadc4, tx=100, y=412, 
tw=3, th=1, color0=1, color1=0, px=0, py=0) at gdevm1.c:493
493	    source_data = tiles->data + ((y + py) % tiles->rep_height) * source_raster;

(gdb) print *tiles
$4 = {
  data = 0x0, 
  raster = 0, 
  size = {
    x = 0, 
    y = 0
  }, 
  id = 0, 
  rep_width = 0, 
  rep_height = 0, 
  rep_shift = 0, 
  shift = 0
}

Since this only occurs when using banding, assigning to Igor.

This is a regression, it worked in gs8.54 but not in gs8.56 and later.  The revision which broke this is 
r7861.

Bumping priority for crash.

Created attachment 4228 [details]
imagemask-with-pattern.ps

The failure happens because the clist language do not pass tile id. In same 
time, the implementation of strip_tile_rectangle for clist writer 
(clist_strip_tile_rectangle) assumes that a tile id presents whenever the tile 
is not empty. In the test case a clist based pattern includes a tile, so while 
its playback to the page clist device clist_strip_tile_rectangle recieves a 
non-empty tile with zero tile id and then malfunctions.

So possible resolutions are either to extend the clist language with tile id 
(see clist_change_tile), or fall back to gx_default_strip_tile_rectangle from 
clist_strip_tile_rectangle when dev->proc.open_device==clist_open. I'm not 
clear which way is better.

I wonder that neither my local test base, nor 
incasper:/home/regression/comparefiles include imagemask-with-pattern.ps . I 
see it comes from the closed bug 689210, but why it is not tested nightly ?

The nightly regression test runs on peeves, and there is a
/home/regression/comparefiles/imagemask-with-pattern.ps

It's also in the ghostcript svn tree; under tests/ps

I think clist_change_tile must check for cldev->nbands == 1. Well it's a new 
invariant, which allows to know whethet the target is a clist writer for a 
clist-based pattern accumulator. If true, write tiles->id to 
cmd_opv_set_tile_size. Functions cmd_size_tile_params, cmd_store_tile_params, 
read_set_tile_size will need an extra boolean parameter for passing the 
condition mentioned above. 

Regarding Comment #8 : cdev->nbands==1 doewsn't work properly because it 
appears so when running -sDEVICE=pbmraw -dMaxBitmap=10000 CompareFiles\000-
00.ps . 

A right criterion is :

#define IS_CLIST_FOR_PATTERN(cdev) (cdev->procs.open_device == 
pattern_clist_open_device)

It could break non-pattern builds. Not sure whether it is important.

The case where nbands is 1 does occur when -dMaxBitmap is small enough to
force banding, but -dBufferSpace is large enough that a given page only needs
1 band.

We have at least one customer that always wants banding, so they specify
-dMaxBitmap=10000 always.

Created attachment 4230 [details]
patch.txt

A suggested patch waits for code review from Henry and Ray.

Patch to HEAD :
http://ghostscript.com/pipermail/gs-cvs/2008-July/008440.html
Note it was committed with a wrong lig message that was fixed later.

This is still broken, at least on Mac OS X and Linux.  On both my MacBook Pro and on peeves the 
following command line still gives a seg fault with gshead (r8865):

  bin/gs -sDEVICE=pbmraw -o test.pbm -r600 -dMaxBitmap=10000 ./imagemask-with-pattern.ps

Here is the output:

marcos@peeves:[27]% bin/gs  -sDEVICE=pbmraw -o test.pbm -r600 -dMaxBitmap=10000 
./imagemask-with-pattern.ps
GPL Ghostscript SVN PRE-RELEASE 8.63 (2008-03-01)
Copyright (C) 2008 Artifex Software, Inc.  All rights reserved.
This software comes with NO WARRANTY: see the file PUBLIC for details.
Floating exception
marcos@peeves:[28]% 

Well Comment #13 supplies a different command line than ones above. I could 
reproduce the crash on Windows. Will analyze it.

One more patch to HEAD :
http://ghostscript.com/pipermail/gs-cvs/2008-July/008445.html