None of the allocations in ccomps could tolerate failure. So this change makes
them all call wrappers that cleanly abort in the event of out-of-memory. Note
that this also fixes an issue where ccomps incorrectly identified itself as `gc`
in one of the previous failure messages.
The `undef` part of this macro juggling was typoed as `WIN32_STATIC` in commit b26b5fc076c1b5d9919ce79c807f5b3921149597, so it never properly undid the
preceding `inline` redirection. However, this redirection is unnecessary anyway.
Contrary to the Microsoft docs,¹ the `inline` keyword seems understood in *both*
C and C++.
Prior to this commit, the gdefs.h header was generated by a C program, mkdefs.c.
There were a number of issues with this approach:
1. The CMake build system was assuming the compiler to build mkdefs.c and the
compiler to build Graphviz itself were the same. This is not necessarily
true when cross-compiling.
2. Generation under MSBuild seems to have been impractical, so the generated
header was checked in to the repository under windows/include/gvpr/gdefs.h,
somewhat defeating the purpose of making it generated.
3. The CMake build system seems to not have been setup to correctly compile
mkdefs.c under all circumstances (see #2101).
This change removes any reliance on a host C compiler and instead uses a series
of X macros¹ to achieve the same effect. The values of all generated constants
and the content of generated structures is intended to be unchanged, though some
#defines have been altered to enums. In these cases, there was no advantage to
using a macro and multiple advantages to not using a macro.
This change is affecting a shipped header (gdefs.h) and also removes it from the
list of shipped headers. Installing it appears to have been a mistake as there
is no easy way for end users to use it. The header, fully expanded, still relies
on further expansion of macros that are only defined in expr.h, a header that is
not shipped.
lib/expr: use RSH as a constant for >> instead of RS
A couple of places in the code base are dodging an `RS` symbol apparently
introduced by an HPPA header. It is simpler to just use a symbol that does not
collide at all.
Chris Mansley [Wed, 21 Jul 2021 19:33:42 +0000 (19:33 +0000)]
Initialize nPasses in aspect_t struct in SetAspect
In dotLayout, nPasses is decremented in a while loop. Since it may be
uninitialized, this triggers clang's UndefinedBehaviorSanitizer:
signed-integer-overflow. While behavior does not change because the loop also
checks nextIter, which is initialized, this change fixes the sanitizer error.
rewrite edgepaint command line parsing using getopt_long
This change makes edgepaint command line parsing code more standard and robust.
It now rejects invalid arguments and takes more standard GNU-style double dash
prefixed options. Note that the old style single dash prefixed options are also
still accepted for compatibility purposes. Closes #1971.
When passing the option `-lightness` multiple times, pointers to previous
lightness strings would be overwritten and lost. This is unlikely to have had a
significant effect. Related to #1971.
add a test case for edgepaint command line parsing
This is a safe guard for upcoming changes that will rewrite how edgepaint parses
command line options. We want to ensure the existing interpretation of options
is not broken by this change. Related to #1971.
reflow edgepaint man page and make content consistent
Text is now wrapped at 80 columns (in the source, not the `man` output display),
sentences begin with capitals and end with periods, and `\fR` is used instead of
`\fP`. `\fR` sets regular text while `\fP` returns to the previous font. On the
surface, `\fP` sounds better, but troff only remembers the immediately prior
font, rather than a stack. So `\fP`s do not nest. With this in mind, `\fR` is
simpler and more comprehensible.
str_mpy: [nfc] pre-compute and allocate the result string
This change does not affect the functionality of this function, but it has two
motivating advantages:
1. The temporary scratch buffer `ex->tmp` is no longer used. Though it is not
obvious without auditing a lot of surrounding code, the data written into
this buffer does not need to be retained beyond the lifetime of this
function. Removing its use not only removes a code path through sfio, but
decouples this code from other code using `ex->tmp` making it easier to
understand. Related to #1873, #1998.
2. The prior code used an sfio temporary buffer to construct the result string
and then duplicated it into a vmalloc-allocated buffer. This is reasonable
as vmalloc has no support for incrementally constructing dynamically
allocated strings. However we can avoid the intermediate sfio buffer by
simply pre-computing the final vmalloc allocation that will be needed. This
change does exactly that and simply writes the result once into its final
destination instead of copying through an intermediate buffer. This
should not only (slightly) decrease transient heap pressure, but also
(again slightly) accelerate the performance of this function.
Both these effects are a simplification with respect to how the compiler sees
this function. That is, an optimizing compiler should now better comprehend the
intent of this function and be able to more aggressively specialize and inline
it where relevant.
Upcoming changes will improve the efficiency of this function and decrease its
coupling with other operations. Rather than introduce these new changes in a
differing style, this preparatory commit rewrites the existing functionality in
this style first, without affecting its behavior. Related to #1873, #1998.
str_mod: [nfc] pre-compute and allocate the result string
This change does not affect the functionality of this function, but it has two
motivating advantages:
1. The temporary scratch buffer `ex->tmp` is no longer used. Though it is not
obvious without auditing a lot of surrounding code, the data written into
this buffer does not need to be retained beyond the lifetime of this
function. Removing its use not only removes a code path through sfio, but
decouples this code from other code using `ex->tmp` making it easier to
understand. Related to #1873, #1998.
2. The prior code used an sfio temporary buffer to construct the result string
and then duplicated it into a vmalloc-allocated buffer. This is reasonable
as vmalloc has no support for incrementally constructing dynamically
allocated strings. However we can avoid the intermediate sfio buffer by
simply pre-computing the final vmalloc allocation that will be needed. This
change does exactly that and simply writes the result once into its final
destination instead of copying through an intermediate buffer. This
should not only (slightly) decrease transient heap pressure, but also
(again slightly) accelerate the performance of this function.
Both these effects are a simplification with respect to how the compiler sees
this function. That is, an optimizing compiler should now better comprehend the
intent of this function and be able to more aggressively specialize and inline
it where relevant.
Upcoming changes will improve the efficiency of this function and decrease its
coupling with other operations. Rather than introduce these new changes in a
differing style, this preparatory commit rewrites the existing functionality in
this style first, without affecting its behavior. Related to #1873, #1998.
str_xor: [nfc] pre-compute and allocate the result string
This change does not affect the functionality of this function, but it has two
motivating advantages:
1. The temporary scratch buffer `ex->tmp` is no longer used. Though it is not
obvious without auditing a lot of surrounding code, the data written into
this buffer does not need to be retained beyond the lifetime of this
function. Removing its use not only removes a code path through sfio, but
decouples this code from other code using `ex->tmp` making it easier to
understand. Related to #1873, #1998.
2. The prior code used an sfio temporary buffer to construct the result string
and then duplicated it into a vmalloc-allocated buffer. This is reasonable
as vmalloc has no support for incrementally constructing dynamically
allocated strings. However we can avoid the intermediate sfio buffer by
simply pre-computing the final vmalloc allocation that will be needed. This
change does exactly that and simply writes the result once into its final
destination instead of copying through an intermediate buffer. This
should not only (slightly) decrease transient heap pressure, but also
(again slightly) accelerate the performance of this function.
Both these effects are a simplification with respect to how the compiler sees
this function. That is, an optimizing compiler should now better comprehend the
intent of this function and be able to more aggressively specialize and inline
it where relevant.
Upcoming changes will improve the efficiency of this function and decrease its
coupling with other operations. Rather than introduce these new changes in a
differing style, this preparatory commit rewrites the existing functionality in
this style first, without affecting its behavior. Related to #1873, #1998.
str_and: [nfc] pre-compute and allocate the result string
This change does not affect the functionality of this function, but it has two
motivating advantages:
1. The temporary scratch buffer `ex->tmp` is no longer used. Though it is not
obvious without auditing a lot of surrounding code, the data written into
this buffer does not need to be retained beyond the lifetime of this
function. Removing its use not only removes a code path through sfio, but
decouples this code from other code using `ex->tmp` making it easier to
understand. Related to #1873, #1998.
2. The prior code used an sfio temporary buffer to construct the result string
and then duplicated it into a vmalloc-allocated buffer. This is reasonable
as vmalloc has no support for incrementally constructing dynamically
allocated strings. However we can avoid the intermediate sfio buffer by
simply pre-computing the final vmalloc allocation that will be needed. This
change does exactly that and simply writes the result once into its final
destination instead of copying through an intermediate buffer. This
should not only (slightly) decrease transient heap pressure, but also
(again slightly) accelerate the performance of this function.
Both these effects are a simplification with respect to how the compiler sees
this function. That is, an optimizing compiler should now better comprehend the
intent of this function and be able to more aggressively specialize and inline
it where relevant.
Upcoming changes will improve the efficiency of this function and decrease its
coupling with other operations. Rather than introduce these new changes in a
differing style, this preparatory commit rewrites the existing functionality in
this style first, without affecting its behavior. Related to #1873, #1998.
str_ior: [nfc] pre-compute and allocate the result string
This change does not affect the functionality of this function, but it has two
motivating advantages:
1. The temporary scratch buffer `ex->tmp` is no longer used. Though it is not
obvious without auditing a lot of surrounding code, the data written into
this buffer does not need to be retained beyond the lifetime of this
function. Removing its use not only removes a code path through sfio, but
decouples this code from other code using `ex->tmp` making it easier to
understand. Related to #1873, #1998.
2. The prior code used an sfio temporary buffer to construct the result string
and then duplicated it into a vmalloc-allocated buffer. This is reasonable
as vmalloc has no support for incrementally constructing dynamically
allocated strings. However we can avoid the intermediate sfio buffer by
simply pre-computing the final vmalloc allocation that will be needed. This
change does exactly that and simply writes the result once into its final
destination instead of copying through an intermediate buffer. This
should not only (slightly) decrease transient heap pressure, but also
(again slightly) accelerate the performance of this function.
Both these effects are a simplification with respect to how the compiler sees
this function. That is, an optimizing compiler should now better comprehend the
intent of this function and be able to more aggressively specialize and inline
it where relevant.
Upcoming changes will improve the efficiency of this function and decrease its
coupling with other operations. Rather than introduce these new changes in a
differing style, this preparatory commit rewrites the existing functionality in
this style first, without affecting its behavior. Related to #1873, #1998.
agcallbacks: force flag to be treated as a boolean
We cannot easily change this function’s signature to take a C99 bool without
breaking API, but we can at least ensure it is treating its flag as a boolean
internally. This squashes two -Wconversion warnings.
Now that the switched-on type is represented as an enum, it is simpler to audit
all call sites and confirm that the default case of all these switches is
unreachable.
Nothing was relying on the values of these constants and it is clearer to use
an enum here. It is now more obvious to readers when one of these values is
being passed between functions instead of an ambiguous int.
remove vmalloc indirection through function pointers
In the past, vmalloc supported a configurable allocation method. That is, it was
parametrized with the underlying allocator. Commit 099964281cfccd7b4b0dbc4473ad9347ed3ca2f0 removed support for this and made the
so-called “bestalloc” the only option, directly using malloc. Since then, the
Vmethod_t part of vmalloc has been more-or-less pure overhead. Allocator
function pointers were maintained here and all calls were indirected through
these function pointers.
This commit removes these function pointers (reducing allocation of the
Vmalloc_t struct by sizeof(void*) * 3) and exposes the (only) implementation
behind these pointers as direct functions that can be called. It is expected
this will accelerate the performance of any lib/expr code heavily using vmalloc.
Apart from performance impact, this is a significant simplification of the
vmalloc code.
This change has no effect on functionality, but strchr is cheaper to call and
equivalent to these strstr calls. This likely makes no difference in an
optimized build as modern compilers can see this transformation is possible
themselves. However, this change may assist older compilers or accelerate
unoptimized builds.
fix: remove hard limit of 1000 boxes in dot spline code
The dot spline code used a static array of 1000 box data structures. It is not
clear to me why this limit was thought to be enough. I suspect the limit (added
in the initial Graphviz revision) was just thought to be something sufficiently
large that no user would ever hit it.
This is no longer true. The graph in issue #2095 exceeds this limit. Because
there are no bounds checks when moving through the boxes array, this resulted in
a crash due to a write beyond the end of the array.
In this commit, we remove this limit entirely and instead use a dynamically
allocated expanding array of boxes. This permits handling an arbitrary number of
boxes during computation. Fixes #2095.
make_flat_edge: use a local box array instead of the global boxes
The values written to this array during make_flat_edge do not need to be
retained after the function returns. It was simply using the boxes as scratch
space. Using a local array instead makes it easier for the compiler to optimize
and will ease some upcoming changes. Related to #2095.
This introduces a new -Wshadow compiler warning, but that will be removed in a
future commit.
make_flat_bottom_edges: use a local box array instead of the global boxes
The values written to this array during make_flat_bottom_edges do not need to be
retained after the function returns. It was simply using the boxes as scratch
space. Using a local array instead makes it easier for the compiler to optimize
and will ease some upcoming changes. Related to #2095.
This introduces a new -Wshadow compiler warning, but that will be removed in a
future commit.
make_flat_labeled_edge: use a local box array instead of the global boxes
The values written to this array during make_flat_labeled_edge do not need to be
retained after the function returns. It was simply using the boxes as scratch
space. Using a local array instead makes it easier for the compiler to optimize
and will ease some upcoming changes. Related to #2095.
This introduces a new -Wshadow compiler warning, but that will be removed in a
future commit.
As I look into how to resolve this issue, I realize it does not make sense to
look for mention of a subgraph because the layout attribute is invalid on *all*
entities except graphs. That is, the error/warning should just say it is only
valid for graphs. Related to #2078.
rewrite streq as a function and remove micro-optimization
There is no need for this to be a macro or for it to check the first character
explicitly. Modern compilers can do this kind of optimization themselves.
I do not know why the comment says it is surprising the assertion fails. It
seems perfectly normal to me that it would fail, as the correct condition is
against *both* AGINEDGE and AGOUTEDGE as in the following line.
projects / graphviz / log