

	***** OK, why write another toy C compiler?

	It's better than getting drunk in a bar.


	***** Why doesn't signal() work on some systems?

	The signal(2) system call exists only on rather ancient Unix
	systems. Modern systems implement it in libc as a thin wrapper
	around the sigaction(2) system call. This is what SubC's libc
	also attempts to do.

	However, in the latest iteration of Unix, the sigaction(2)
	system call itself seems to be a wrapper around something
	even more complex. Basically a lot of the signal handling
	functionality is implemented in libc on those systems, and
	installing signal handlers with sigaction(2) in the usual
	way results in core dumps when invoking a signal handler.

	This is because "signal trampolines" have moved from kernel
	space to user space and installing the proper trampoline is
	now done on a per-signal basis by some code in libc. See, for
	example, the sigtramp(9) man page on NetBSD for all the gory
	details.

	Unfortunately, K&R II specifies signal() as a standard function,
	so I am doing my best to implement it for all supported systems.
	It seems to work fine on most 32-bit Unixes and seems to crash
	on most 64-bit ones, except for FreeBSD (10.1-RELEASE).

	Note that SIG_DFL and SIG_IGN do work, even on newer systems,
	only signal handlers crash when invoked.

	The only work-around is to use the signal wrapper in the
	system's C library, which may cause other problems, though.
	It also causes a massive increase in executable size.


	***** What if the triple test fails?

	It depends. When the test script reports that the triple test
	failed, it failed, and you cannot assume that you have a
	working compiler. In this case, *please* send a bug report to
	me and include the _stagedump.tar file that the test script
	generated.

	The test may also pass at assembly level, but then fail at
	binary level, i.e. stages 2 and 3 generate identical assembly
	programs, but the binaries differ nevertheless. Even in this
	case, please send me the file generated by the script!

	When the compiler passes at assembly level, you can assume that
	it works fine and the cause of failure is beyond its scope. The
	most probably cause is that the executables on your system
	contain a timestamp or some other kind of build id.

	However, I have also seen stage 2/3 binaries that differed at
	thousands of locations, while assembly output still matched,
	which is weird. So especially if you get stage 2/3 binaries
	that are completely different (not only in a few bytes), please
	do send me the test script output!


	***** What about code quality?

	Code quality has improved a lot since the 2012 book version:

				  Book  5/2014
		scc -t *.c	 0.53s	 0.48s   (9% faster)
		text size	66,576	52,672  (21% smaller)

	Run times were obtained by syntax-testing and compiling the book
	version on a 386-class machine. The 2014 compiler is 9% faster,
	although it uses an optimizer, and generates 21% smaller code.

	SubC generates compact, but not very runtime-efficient code. In
	general, its code reaches about 67% of the speed of code emitted
	by GCC or Clang at optimization level 0.

	The compiler executables generated by SubC are tyically <33% the
	size of those generated by GCC or Clang, though. This is mostly
	because SubC does not use a bloated libc implementation.

	Here are some concrete numbers. Run times were obtained by
	bootstrapping the SubC compiler multiple times and taking the
	best results. All sizes refer to statically linked executables.

	The boot stage 0 time is the time taken by the GCC-compiled
	or Clang-compiled SubC compilers, while the boot stage 1 time
	is the time taken by a self-compiled SubC compiler.

	FreeBSD 8.2-RELEASE, 686 CPU @ 600 MHz
				GCC 4.2.1	SUBC 2014-05-25
		boot stage 0	  2.40s		  0.80s
		boot stage 1			  1.35s (40% slower)
		text size	229,965		 71,764 (75% smaller)

	FreeBSD 10.0-RC5, armv6 CPU @ 700 MHz
				Clang 3.3	SUBC 2014-05-25
		boot stage 0	 14.22s		  4.68s 
		boot stage 1			  6.14s (24% slower)
		text size	450,411		143,508 (69% smaller)

	FreeBSD 10.0-RELEASE, x86-64 CPU @ 875 MHz
				Clang 3.3	SUBC 2014-05-25
		boot stage 0	  1.26s		  0.67s
		boot stage 1			  1.03s (35% slower)
		text size	478,370		 93,510 (80% smaller)

