1. Wei Hu
  2. HLVM


HLVM / ChangeLog

Sat Jan 16 12:17:05 GMT 2010  Jon Harrop <jon@ffconsultancy.com>

	* Disabling TCO now uses a separate set of tests that do not require
	TCO so they should all run.
	* Applied bluestorm's patches for Makefile and camlp4 macros to make it
	easier to write HLVM IR.

Sun Jan 10 16:29:16 GMT 2010  Jon Harrop <jon@ffconsultancy.com>

	* Removed the use of AddressOf in the hash table test.
	* Added --no-stack-handler option so you can remove the libsigsegv
	handler that is used to detect stack overflows.
	* Convert some remaining uses of GetValue in the GC to use the abstract
	Seq.* API.
	* Fixed a bug where Alloc wasn't initializing mark state so it could
	already be marked when the mark phase starts and its children could go
	unmarked and get reclaimed despite being reachable.

Thu Jan  7 15:18:59 GMT 2010  Jon Harrop <jon@ffconsultancy.com>

	* Added a `Byte type to HLVM to make strings easier to use.
	* Ported the ray tracer benchmark from OCaml to HLVM. HLVM is
	  significantly faster despite the use of the FFI to do sqrt.
	* Made runtime.cpp 64-bit happy (int -> long).

Mon Jan  4 07:49:18 GMT 2010  Jon Harrop <jon@ffconsultancy.com>

	* Inlined get_stack.
	* Changed the value "unit" to an undef rather than "int 0".
	* Fixed unnecessary allocation of mark state for empty (NULL) arrays:
	parallel n-queens is ~3% faster.
	* Fixed the calculator example.

Fri Jan  1 20:57:09 GMT 2010  Jon Harrop <jon@ffconsultancy.com>

	* Optimized thread-local data to be passed by pointer in function
	arguments, dramatically improving performance.
	* Removed some concurrency bugs in the GC.
	* Altered the parallel n-queens test to recursively spawn child
	* Implemented ptr->size map in runtime.cpp to check for double frees
	when runtime is in debugging mode.

Mon Nov 30 01:14:49 GMT 2009  Jon Harrop <jon@ffconsultancy.com>

	* First working version of threads and stop-the-world GC. Only took 5
	days to implement!
	* Global list of active threads and GC state (running or stopping the
	* Thread local state keeps thread state (running or suspended) and
	shadow stack.
	* First thread to hit allocation quota starts a GC by suspending all
	other threads and performing the same mark+sweep algorithm, albeit
	using all shadow stacks, before resuming all other threads. The other
	threads spin while the GC is running.
	* Currently, every function checks the global state on entry to see if
	it needs to suspend itself. This is contention for the same lock from
	every function call on every thread which is very wasteful and is the
	main reason why HLVM is now 3x slower than it was. However, it is
	easily remedied by removing checks from leaf functions (that make no
	calls) and using a counter or even measuring elapsed real time to
	check less often in non-leaf functions.
	* Only non-trivial threading aspect is to enter and leave blocking
	sections when users perform blocking operations like taking out locks
	of their own. This ensures that one thread cannot start a GC cycle with
	a lock taken out while another thread is waiting on that lock, which
	would deadlock all threads.

Thu Nov 12 18:08:41 GMT 2009  Jon Harrop <jon@ffconsultancy.com>

	* Time spent in mark and sweep phases is accumulated per expression
	* Benchmarks are now warmed up to remove JIT time.
	* New bubblesort benchmark.
	* More fine-grained and intelligent loop unrolling.

Sun Oct 25 22:40:03 GMT 2009  Jon Harrop <jon@ffconsultancy.com>

	* Updated to work with LLVM 2.6: now uses llcontext and initializes
	targets before creating a JIT.
	* Replaced the garbage collector with a more conventional mark-sweep
	collector that stores mark state in the heap.

Sat Sep 26 22:51:59 BST 2009  Jon Harrop <jon@ffconsultancy.com>

        * Array allocation now accepts the initial element value.
        * The second iteration of the compiler is now in SVN.

Thu Jun 25 02:36:38 BST 2009  Jon Harrop <jon@ffconsultancy.com>

	* Improved parsing in the example compiler to handle line numbers and
	* Improved pattern matching in the example compiler to allow tuples to
	be deconstructed in "let .. in" bindings as well as function

Sun Jun 21 20:04:48 BST 2009  Jon Harrop <jon@ffconsultancy.com>

	* Added two examples: a trivial calculator and a rudimentary compiler
	with support for recursive function definitions with explicit types and
	unit, char, int, float and tuple types. Arrays will require
	* Note: closures that do not require cycles can be represented entirely

Sat Mar 21 17:42:31 GMT 2009  Jon Harrop <jon@ffconsultancy.com>

	* Memoized constant string literals.
	* Fixed perf bug where all bound variables were being pushed onto the
	shadow stack at the beginning of every function body instead of just
	the function's arguments.
	* Optimized array visit function to skip arrays that do not contain
	any reference types.
	* Implemented standalone compilation. All JIT executed functions are
	called from a "main" function in the output aout.bc which the ./opt.sh
	script compiles through LLVM's optimization passes into aoutopt. This
	gives us some idea of how much performance would be gained by applying
	LLVM's optimization passes if we had bindings for them.
	* Moved timing code from the compiler into the generated code so our
	standalone executable can be used for benchmarking.

Fri Mar 13 02:42:33 GMT 2009  Jon Harrop <jon@ffconsultancy.com>

	* Switched from LLVM's malloc intrinsic to helper functions for alloc
	(using calloc in order to initialize references to NULL) and free.
	* Optimized the GC to use a hash table instead of linear search,
	improving performance by up to 100x on our benchmarks.
	* Avoided resetting the shadow stack at the ends of functions that do
	not push any roots onto it.

Wed Mar 11 21:05:44 GMT 2009  Jon Harrop <jon@ffconsultancy.com>

	* Chose to put new LLVM bindings into the HLVM project instead of
	patching LLVM. We now have bindings for BuildExtractValue,
	BuildInsertValue and for enabling tail call optimization globally.

Wed Mar 11 04:01:45 GMT 2009  Jon Harrop <jon@ffconsultancy.com>

	* Tidied the run-time type code to reuse HLVM's own structs.
	* Fixed a perf bug where reading a reference type from a bound
	variable was pushing it onto the shadow stack. Sieve is now 70% faster.
	* Fixed a bug in init_type which was invoking a fastcc function when it
	should have been using the C calling convention.
	* Added dummy arguments to all argumentless invocations to work around
	a bug in LLVM's OCaml bindings, which calls malloc for zero bytes
	(which is illegal) if there are zero arguments.

Mon Mar  9 12:31:16 GMT 2009  Jon Harrop <jon@ffconsultancy.com>

	* Fixed the GC to use an explicit "visit stack" instead of the system
	stack. This eliminates the stack overflow bug that prevented the list
	test from running to completion but the current GC is so slow that
	I still have not run this test to completion.

Sun Mar  8 04:23:30 GMT 2009  Jon Harrop <jon@ffconsultancy.com>

	* Prepared for first public release.

Fri Mar  6 01:56:13 GMT 2009  Jon Harrop <jon@ffconsultancy.com>

	* Made the "expr" and "def" functions mutually recursive so the 
	compilation of an expression can trigger the compilation of an
	auxiliary function that it depends upon.
	* Implemented a simple stop-the-world mark and sweep garbage collector
	written in the HLVM language itself.
	* A new `UnsafeFunction top-level statement defines a function without
	injecting any code for the GC.
	* Ordinary functions call gc_root to register a local root into the
	heap (e.g. when loading a reference type) by pushing it onto the shadow
	stack, call gc_alloc when a value is allocated to push it onto the
	allocated list and gc_restore when a function returns to unwind the
	shadow stack.

Tue Mar  3 02:38:35 GMT 2009  Jon Harrop <jon@ffconsultancy.com>

	* Realised that the current functionality can already be used to
	represent closures as a function pointer and a reference type
	containing the environment. Added some tests that use this.

Wed Feb 25 21:20:26 GMT 2009  Jon Harrop <jon@ffconsultancy.com>

	* Made ffib benchmark 50% faster by swapping the branches of the "if"
	expression: most common branch should go first.
	* Added code to Printf to handle 32-bit floats by extending them to
	64-bit floats for printing.

Wed Feb 25 03:22:42 GMT 2009  Jon Harrop <jon@ffconsultancy.com>

	* Added OCaml bindings to the new ExtractElement and InsertElement
	instructions now in LLVM's C bindings. This has improved performance
	a lot: Mandelbrot3 (with complex arithmetic) is now 3.5x faster and
	HLVM is faster than OCaml on every test.

Sun Feb 22 16:14:10 GMT 2009  Jon Harrop <jon@ffconsultancy.com>

	* First implementation of reference types represented as a struct
	containing a pointer to the type and a pointer to the argument (if
	any). This allows typed NULL references which should permit efficient
	implementation of properly-typed linked lists.
	* Returning first-class structs inhibits TCO so we must sret all struct
	return types, which I have now done.
	* Removed a layer of indirection in the representation of arrays, which
	is now a struct containing the type, length and pointer to heap data.
	* Fixed with respect to all existing tests. Array performance has
	degraded slightly due to the current overhead of structs.

Thu Feb 19 23:23:25 GMT 2009  Jon Harrop <jon@ffconsultancy.com>

	* build_allocas must go in the entry block of a function. Altered
	state#alloca to insert at the end of the entry block and tweaked the
	compilation of "if" expressions to insert the final jump from the
	previous block only after compilation of both branches is complete, so
	they insert their allocas before the branch.
	* Lots of optimization passes are not yet available from OCaml but
	they could do a great deal with our code, particularly our inefficient
	struct code.
	* Reference type is a struct with a pointer to the type and a pointer
	to the argument. That allows NULL to be typed.

Fri Feb  6 01:39:13 GMT 2009  Jon Harrop <jon@ffconsultancy.com>

	* Added structs but they are locally allocated so they cannot be
	passed in the arguments to a tail call.
	* Used an ingenious workaround: copy every struct into an alloca in
	order to use gep until extractvalue is implemented.
	* Fixed a bug caused by marking every call as a tail call when this
	can confuse LLVM's (CSE) optimization pass.

Tue Jan 27 00:47:10 GMT 2009  Jon Harrop <jon@ffconsultancy.com>

	* Extended sieve and mandelbrot examples to illustrate segmentation
	fault due to stack overflow due to non-tail calls.
	* Introduced a Return constructor into the Expr.t type to handle tail
	calls properly.
	* Added first-class function pointers by casting all defined functions
	into pointer types before binding them to their variable name.

Sun Jan 25 06:51:00 GMT 2009  Jon Harrop <jon@ffconsultancy.com>

	* Clean design with "expr" checking all types including return types.
	* FastCC but no propagation of tail calls into final expressions.
	* No extern functions.
	* No first-class function pointers.
	* No run-time types.
	* No shadow stack.
	* No tuples or multiple return values: functions accept multiple
	arguments but return only one result.