Conflicts:
	capn-malloc.c

Fixes a compiler warning.

Renamed gtest/include to gtest/gtest since that's what the include files
expected.

Adjusted the Makefile to work through it all.

No real need to build a gtest archive, gtest-all.cc includes all the
files so a single .o is fine.

It is not used and we already validated the total buffer size
beforehand.

Should make it a bit more readable, the variable is used before that
case as well but doesn't require the zero initialization in that case.

The hdr variable gets the same constants as value but it is easier to
understand that the constants are used than the hdr which needs to be
traced to the value it received.

The header of packing is a single byte and nothing was written to the
first byte anyway.

This uses the capn_deflate function.

The current use may get smaller packing than the maximum possible due to
the inability to peak around segments as it packs each segment on its
own. This saves time compared to copying everything into one place and
saves effort by not requiring to change the interface of capn_deflate to
support multiple different buffers.

It should be possible to make the capn_deflate state machine better to
handle the multiple buffers case too.

* Rip out the gtest built in here.  Use the system library.
* Test against gtest-1.7.0-1 on Arch Linux
* Add the missing -lpthread library

* These should be easily verified as correct.
* Tested on gcc-5.2.0 and clang-3.6.2

* Never referenced and generate warnings on modern compilers
* Tested on gcc-5.2.0 and clang-3.6.2

* Later versions of compilers/pre-processors are upset about the syntax
* Bracket unterminated string with proper comment tags

* My compilers mark these as errors
* Attempt to be more correct
* Tested on gcc-5.2.0 and clang-3.6.2

* Return null when appropriate and assign a value earlier.

(also makes the next merge much easier)

(also makes the next merge much easier)

…sent length"

This reverts commit 0b6f952.

(for merge)

Conflicts:
	Makefile		-- manually resolved
	compiler/schema.capnp	-- used from liamstask
	compiler/test.capnp	-- used from liamstask

This reverts commit 9fe7167.

(this makes the following merge easier)

Conflicts: (manually resolved)
	capn-malloc.c
	capn-stream.c
	capn-test.cpp
	capn.c
	capn.h
	compiler/capnpc-c.c

(No changes in 640Labs that are useful to merge -- just mark as merged.)

Travis should autodetect & do the right thing for autoconf projects...

... it doesn't autodetect that it needs to run autoreconf
Also add gtest package.

automake 1.11 is sufficient (and what's installed by default on CI
boxes).  Also disable cap'n proto reference (it's not actually used)

gtest packages intentionally don't provide binaries anymore.
Will probably re-add as submodule later.

Pointers are written as offsets to the segment start;  leaving p->data
as NULL results in invalid pointers for zero-size objects (particularly,
lists.)

capn_new_string is exported as symbol but not in the header
capn_write_mem_packed is internal but the symbol is visible

"struct { } foo" is not valid ISO C99; while gcc/clang/icc support it,
other compilers and tools don't (e.g. pycparser)

Add build instructions

ssize_t is not available in visual studio, nor is sys/param.h We typedef
intmax_t to ssize_t

Visual studio's execution flow analysis insists that (e) could be
undefined. I don't see it, but this permits compilation.

Unless capn_segment is defined with __declspec(align(64)),
check_segment_alignment fails to compile in x86 mode, as (sizeof(struct
capn_segment)&7) -> (44 & 7) evaluates to 4
Always compiles in x64 mode, as (sizeof(struct capn_segment)&7) -> (80 &
7) evaluates to 0

start making endianness switch less weird

working with stricter compiler flags

neither Windows nor OSX have endian.h

autoconf checking this makes things worse on Windows too, so let's just
try endian.h on "unix" systems.

Strict aliasing prohibits writing to union members of one type, then
reading another type.  The read is undefined.  Use memcpy instead (which
compilers will hopefully optimize away...)

allows changing output filenames from "foo.capnp.c" to
"foo.capnpSOMETHING.c" where SOMETHING is the argument to the
annotation.

init_fp itself tries to read 4 bytes for the segment count.

Missing braces end up scaling the offset by * 4 ... which is actually a
stack/heap overflow.

There is no point in having these functions be part of the public API.

This reverts commit 1afce95.

Turns out this is a misunderstanding of the C standard...  meh.

Fixes github issue #11, also makes it easier to use annotations.

Generating per-field getters/setters adds a lot of functions; this makes
them switchable with an annotation in the source schema.

This fixes #include output when $C.nameinfix is used.

For capn_write_fd() the write() function would be ideal to pass as
write_fd, but write() has type ssize_t (*)(int, const void *, size_t),
whereas write_fd expects ssize_t (*)(int, void *, size_t). Passing
write() directly with GCC 5.4 causes a warning
-Wincompatible-pointer-types (on by default).

…s build on 32 bits arm.