~centralelyon2010/inkscape/imagelinks2

This file contains some loose proofs of a few properties.  It's somewhat
ad-hoc.  At least it gives an indication of what assert/g_assert calls have
been checked by a developer.  If an assertion does trigger, then this file may
help in debugging that assertion failure.

It's currently ordered by caller.

(Re-ordering to avoid forward references in proofs might be a good idea,
though this would in some cases require splitting up the proofs for a routine,
e.g. proving preconditions of g called from f, then proving g's postcondition,
then using that to prove something else in f again.  Furthermore it may not
even be possible to avoid forward references for recursive/looping code.)



src/pencil-context.cpp:fit_and_split

Very loose proof of !sp_curve_empty (pc->red_curve) assertion:
fit_and_split is called successively with its input varying only by appending a point.
For the n_segs > 0 && unsigned(pc->npoints) < G_N_ELEMENTS(pc->p) condition to fail,
we must have at least 3 distinct points, which means that a previous call had 2 distinct points,
in which case we'd have filled in pc->red_curve to a non-empty curve.

Expansion of the above claim of at least 3 distinct points: We know n_segs <= 0 ||
unsigned(dc->npoints) >= G_N_ELEMENTS(pc->p) from the negation of the containing `if' condition.
G_N_ELEMENTS(pc->p) is greater than 3 (in int arithmetic), from SPPencilContext::p array definition
in pencil-context.h.  npoints grows by no more than one per fit_and_split invocation; we should be
able to establish that dc->npoints == G_N_ELEMENTS(pc->p) if unsigned(dc->npoints) >=
G_N_ELEMENTS(pc->p), in which case 3 <= dc->npoints in int arithmetic.  We know that dc->npoints >=
2 from assertion at head of fit_and_split; in which case if n_segs <= 0 then fit_and_split has
failed, which implies that dc->npoints > 2 (since the fitter can always exactly fit 2 points,
i.e. it never fails if npoints == 2; TODO: add sp_bezier_fit_cubic postcondition for this).


src/pencil-context.cpp:fit_and_split

Proof of precondition: The only caller is spdc_add_freehand_point (by
textual search in that file, and staticness).  See proof for that
function.


src/pencil-context.cpp:spdc_add_freehand_point

Proof of fit_and_split `pc->npoints > 1' requirement:
It initially unconditionally asserts `pc->npoints > 0'.  There are no function calls or modifications
of pc or pc->npoints other than incrementing pc->npoints after that assertion.
We assume that integer overflow doesn't occur during that increment,
so we get pc->npoints > 1.


src/pencil-context.cpp:spdc_set_endpoint

Very loose proof of npoints > 0: Should be preceded by spdc_set_startpoint(pc) according to state
transitions.  spdc_set_startpoint sets pc->npoints to 0 (handled at beginning of function) or 1.


src/display/bezier-utils.cpp:compute_max_error

Proof of postcondition: *splitPoint is set only from i, which goes from 1 to less than last.
i isn't written to in the loop body: only uses are indexing built-in arrays d and u
(and RHS of assignment).


src/display/bezier-utils.cpp:sp_bezier_fit_cubic_full

Proof of `nsegs1 != 0' assertion: nsegs1 is const.  Have already
returned in the (nsegs1 < 0) case, so !(nsegs1 < 0), i.e. nsegs1 >= 0
(given that nsegs1 is gint).  nsegs1 is set to
sp_bezier_fit_cubic_full(_, _, _, splitPoint + 1, ...).  We will show
that sp_bezier_fit_cubic_full ensures len < 2 || ret != 0.  splitPoint
> 0 from compute_max_error postcondition combined with error >=
precondition and thus having handled the compute_max_error returning 0
case: if returned 0 for maxError then maxError <= error * 9.0 would be
true, and we recalculate splitPoint; if the renewed maxError is 0 then
the maxError <= error test will succeed and we return.  If we don't
return, then the most recent compute_max_error must have returned
non-zero, which implies (through compute_max_error postcondition) that
splitPoint would have been set s.t. 0 < splitPoint.  splitPoint is not
subsequently written to.  (It is subsequently passed only to
sp_darray_center_tangent 2nd arg, which is a plain unsigned rather
than reference.)  0 < splitPoint < last guarantees splitPoint + 1 >=
2.  (We know splitPoint + 1 won't overflow because last = len - 1 and
len is of the same type as splitPoint.)  Passing splitPoint + 1 for
len of the call that sets nsegs1 ensures that nsegs1 is non-zero (from
the len < 2 || ret != 0 property that we show below).  QED.

Proof that len < 2 || (failed no-dups precondition) || ret != 0: All
returns are either -1, 0, 1, or nsegs1 + nsegs2.  There are two
literal 0 cases: one conditional on len < 2, and the other for failed
precondition (non-uniqued data).  For the nsegs1 + nsegs2 case, we've
already ruled out nsegs1 < 0 (through conditional return) and nsegs2 <
0 (same).  The nsegs1 + nsegs2 case occurs only when we recurse; we've
already shown the desired property for non-recursive case.  In the
nsegs1 non-recursive case, we have that nsegs1 != 0, which combined
with !(nsegs1 < 0) and !(nsegs2 < 0) implies that nsegs1 + nsegs2
either overflows or is greater than 0.  We should be able to show that
nsegs1 + nsegs2 < len even with exact arithmetic.  (Very loose proof:
given that len >= 2 (from earlier conditional return), we can fit len
points using len-1 segments even using straight line segments.)
nsegs1 and nsegs2 are the same type as len, and we've shown that
nsegs1 + nsegs2 in exact arithmetic is >= 0 from each operand being
non-negative, so nsegs1 + nsegs2 doesn't overflow.  Thus nsegs1 +
nsegs2 > 0.  Thus we have shown for each return point that either the
return value is -1 or > 0 or occurs when len < 2 or in failure of
no-dups precondition.  (We should also show that no-dups outer
precondition being met implies it being met for inner instances of
sp_bezier_fit_cubic_full, because we pass a subsequence of the data
array, and don't write to that array.)  QED.

We should also show that the recursion is finite for the inductive
proof to hold.  The finiteness comes from inner calls having len > 0
and len less than that of outer calls (from splitPoint calculation and
0 < splitPoint < last for recursive case and last < len and transitive
property of < for gint).  If len < 2 then we don't recurse
(conditional return).

We should go over this proof to make it clear that there are no
"forward references" other than for recursive case.  We could also be
more formal in use of inductive proof (e.g. clearly stating what the
base and inductive premises are; namely the non-recursing and
recursing cases of sp_bezier_fit_cubic_full).

Better proof sketch that nseg1 + nsegs2 < len: ret < len for each
recursive case other than where len > 0 precondition fails.  nsegs1 is
calculated for inner len=splitPoint + 1, nsegs2 for inner len=len -
splitPoint.  Assuming exact arithmetic we'll transform that to ret <=
len - 1.  Implies that in exact arithmetic, nsegs1 + nsegs2 <=
(splitPoint + 1 - 1) + (len - splitPoint - 1).  Simplifying RHS (using
exact arithmetic): nsegs1 + nsegs2 <= len - 1, i.e. nsegs1 + nsegs2 <
len.  Presumably we can show that machine arithmetic gets the same
results as exact arithmetic from similar arguments used so far for
showing that overflow doesn't occur.  For the recursive case the
return values are either nsegs1 + nsegs2 or -1.

We should also show that inner preconditions hold, especially the len
> 0 precondition.  (For nsegs1 call, we use 0 < splitPoint and that
splitPoint + 1 doesn't overflow.  For nsegs2 call, we pass len -
splitPoint; combine with splitPoint < last, last = len - 1, and no
overflow.)  We've already sketched a proof for no-dups precondition.
The others are fairly simple.

For error >= 0: error is const, and we pass it to all recursions.

For inner max_beziers >= 1: recursions are conditional on outer
1 < max_beziers before setting rec_max_beziers1 to max_beziers - 1,
and passing rec_max_beziers1 as inner max_beziers value,
so we have outer max_beziers >= 2 so inner max_beziers >= 1.
max_beziers and rec_max_beziers1 are both const.


src/display/bezier-utils.cpp:sp_darray_right_tangent(Point const[], unsigned)

Proof of unit_vector precondition that a != Point(0, 0): our unequal precondition.

Loose (incorrect) proof of unit_vector precondition that neither
coordinate is NaN: our in_svg_plane precondition, and fact that
in_svg_plane returns false if either argument is infinite.  HOWEVER,
the unchecked in_svg_plane precondition isn't currently guaranteed, so
we're just relying on the input points never being infinity (which
might occur with strange zoom settings).


src/display/bezier-utils.cpp:sp_darray_right_tangent(Point const[], unsigned, double)

Loose proof of unit_vector precondition that a != Point(0, 0) for first call to unit_vector:

We've asserted that 0 <= tolerance_sq; combine with tolerance_sq <
distsq and transitivity of <=/< show that 0 < distsq.  Definition of
dot should give us that t != 0.0, given that 0.0 * 0.0 == +0.0, and 0
< +0.0 is false.

Loose proof for the second unit_vector invocation: distsq != 0 from ?:
condition, which should give us that t != Point(0, 0) in the same way
as in the above proof.

Proof of sp_darray_right_tangent(Point[], unsigned) preconditions: We
have the same preconditions, and pass the same arguments.  d, *d and
len are const.



src/extension/internal/ps.cpp:PrintPS::print_fill_style:

Proof of the
   g_return_if_fail( style->fill.type == SP_PAINT_TYPE_COLOR
                     || ( style->fill.type == SP_PAINT_TYPE_PAINTSERVER
                          && SP_IS_GRADIENT(SP_STYLE_FILL_SERVER(style)) ) )
at beginning of function:

rgrep print_fill_style reveals no callers in any other files.  There are two calls in ps.cpp, both
inside an `if' test of that same condition (with no relevant lines between the test and the call).
Each call uses `style' as its second argument, and `style' in print_fill_style refers to its second
parameter.  In both caller & callee, `style' is a const pointer to const, and there is very little
code between the two tests, so the relevant values are very unlikely to change between the two
tests.


Proof of
    g_assert( style->fill.type == SP_PAINT_TYPE_PAINTSERVER
              && SP_IS_GRADIENT(SP_STYLE_FILL_SERVER(style)) ) :

The g_return_if_fail(style->fill.type == SP_PAINT_TYPE_COLOR
                     || ( style->fill.type == SP_PAINT_TYPE_PAINTSERVER
                          && SP_IS_GRADIENT(SP_STYLE_FILL_SERVER(style)) ) )
call at the beginning of the function, and we're in the `else' branch of a test for
style->fill.type == SP_PAINT_TYPE_COLOR, and style is a const pointer to const, so it's likely that
style->fill and the gradient object have the same values throughout.



src/extensions/internal/ps.cpp:PrintPS::fill:

Proof of the two assertions
  g_assert( style->fill.type == SP_PAINT_TYPE_PAINTSERVER
            && SP_IS_GRADIENT(SP_STYLE_FILL_SERVER(style)) ) :

Each is in the `else' branch of a test for `style->fill.type == SP_PAINT_TYPE_COLOR',
within a test for
  ( style->fill.type == SP_PAINT_TYPE_COLOR
    || ( style->fill.type == SP_PAINT_TYPE_PAINTSERVER
         && SP_IS_GRADIENT(SP_STYLE_FILL_SERVER(style)) )  ).

`style' is a const pointer to const, so the values are unlikely to have changed between the tests.



src/seltrans.cpp:sp_sel_trans_update_handles

Proof of requirements of sp_show_handles:

sp_show_handles requirements: !arg1.empty.

Before any call to sp_show_handles is a test `if (... || seltrans.empty) { ...; return; }'
(with no `break' etc. call preventing that `return').
Each subsequent sp_show_handles call uses seltrans as arg1.
seltrans is a reference.  There are no calls between that failing seltrans.empty test
and the sp_show_handles calls that pass seltrans.  The sole call is sp_remove_handles,
probably doesn't have access to seltrans.



src/seltrans.cpp:sp_show_handles

Proof of precondition:

sp_show_handles is static.  Searching reveals calls only in sp_sel_trans_update_handles (proof above).



src/sp-spiral.cpp:sp_spiral_fit_and_draw

Proof of postcondition is_unit_vector(*hat2):

hat2 is set by sp_spiral_get_tangent unconditionally, which Ensures is_unit_vector(*hat2).
We then negate *hat2, which doesn't affect its length.
We pass it only to sp_bezier_fit_cubic_full, which claims constness of *hat2.

Proof of unconditionalness: Not inside if/for/while.  No previous `return'.


src/sp-spiral.cpp:sp_spiral_set_shape

Loose proof of requirements for sp_spiral_fit_and_draw:

 Proof of dstep > 0:
 
  SAMPLE_STEP equals .25.
  spiral->revo is bounded to [0.05, 20.0] (and non-NaN) by various CLAMP calls.
  (TODO: Add precondition, given that those CLAMP calls are outside of this function.)
  SAMPLE_SIZE equals 8.
  dstep is const and equals SAMPLE_STEP / spiral->revo / (SAMPLE_SIZE - 1),
  == 1 / (4 * [.05, 20.0] * 7)
  == 1 / [1.4, 560]
  dstep in [.0018, .714].

 Proof of is_unit_vector(hat1):

  Initially guaranteed by sp_spiral_get_tangent Ensures.
  For subsequent calls, hat1 is set from negated hat2 as set by sp_spiral_fit_and_draw,
  which Ensures is_unit_vector(hat2).



src/style.cpp:sp_css_attr_from_style:

Proof of sp_style_write_string pre `style != NULL':

Passes style as style argument.  style is const, and has already been checked against NULL.


src/style.cpp:sp_css_attr_from_object

Proof of `flags in {IFSET, ALWAYS} precondition:

 $ grep sp_css_attr_from_object `sed 's,#.*,,' make.files `
 file.cpp:  SPCSSAttr *style = sp_css_attr_from_object (SP_DOCUMENT_ROOT (doc));
 selection-chemistry.cpp:  SPCSSAttr *css = sp_css_attr_from_object (SP_OBJECT(item), SP_STYLE_FLAG_ALWAYS);
 selection-chemistry.cpp:  SPCSSAttr *temp = sp_css_attr_from_object (last_element, SP_STYLE_FLAG_IFSET);
 style.cpp:sp_css_attr_from_object (SPObject *object, guint flags)
 style.h:SPCSSAttr *sp_css_attr_from_object(SPObject *object, guint flags = SP_STYLE_FLAG_IFSET);


src/style.cpp:sp_css_attr_from_style

Proof of precondition `style != NULL':

Callers are selection-chemistry.cpp and style.cpp:

 $ grep sp_css_attr_from_style `sed 's,#.*,,' make.files `
selection-chemistry.cpp:  SPCSSAttr *css = sp_css_attr_from_style (query, SP_STYLE_FLAG_ALWAYS);
 style.cpp:sp_css_attr_from_style (SPStyle const *const style, guint flags)
 style.cpp:  return sp_css_attr_from_style (style, flags);
 style.h:SPCSSAttr *sp_css_attr_from_style (SPStyle const *const style, guint flags);

selection-chemistry.cpp caller: query is initialized from sp_style_new()
(which guarantees non-NULL), and is const.

style.cpp caller: preceded by explicit test for NULL:

 $ grep -B2 sp_css_attr_from_style style.cpp|tail -3
     if (style == NULL)
         return NULL;
     return sp_css_attr_from_style (style, flags);




# Local Variables:
# mode:indented-text
# fill-column:99
# End:
# vim: filetype=text:tabstop=8:encoding=utf-8:textwidth=99 :