Generic functions are convenient but slow. During the development we usually want the full dynamic feature of CLOS. However, when we really need a fast binary and do not need the dynamism, the dynamic dispatch in the generic functions should be statically compiled away.
We propose a MOP-based implementation of fast inlined generic functions dispatched in compile-time. The amount of work required to inline your generic function is minimal.
Empirical analysis showed that the resulting code is up to 10 times faster than the standard generic functions.
Tested on SBCL and CCL.
Thanks to the suggestion from @phmarek,
I decided to add a feature called
invalid-branch which uses impl-specific
feature to enable a static type checking.
Read the doc!
Added an additional usage note.
The example code here is in
First, declare the generic function with
metaclass. This metaclass is a subclass of
standard-generic-function. Therefore, unless you use its special feature,
it acts exactly the same as the normal generic functions.
(defgeneric plus (a b) (:generic-function-class inlined-generic-function))
Define the methods as usual.
(defmethod plus :around ((a number) (b number)) (+ a b) ;; not a meaningful operation... (call-next-method)) (defmethod plus ((a fixnum) (b fixnum)) (+ a b)) (defmethod plus ((a float) (b float)) (+ a b))
Define a function which uses it.
(defun func-using-plus (a b) (plus a b))
At this point the gf is not inlined.
; disassembly for FUNC-USING-PLUS ; Size: 24 bytes. Origin: #x100A75A165 ; 65: 488BD1 MOV RDX, RCX ; no-arg-parsing entry point ; 68: 488BFB MOV RDI, RBX ; 6B: 488B059EFFFFFF MOV RAX, [RIP-98] ; #<FDEFINITION for PLUS> ; 72: B904000000 MOV ECX, 4 ; 77: FF7508 PUSH QWORD PTR [RBP+8] ; 7A: FF6009 JMP QWORD PTR [RAX+9]
Now its time to inline the gf. There’s nothing different from inlining a normal function.
In order to inline the generic function, just declare it
inline when you use it.
(defun func-using-plus (a b) (declare (inline plus)) (plus a b)) ; disassembly for FUNC-USING-INLINED-PLUS ; Size: 323 bytes. Origin: #x1002C3BD45 ; D45: 8D41F1 LEA EAX, [RCX-15] ; no-arg-parsing entry point ; D48: A80F TEST AL, 15 ; D4A: 755F JNE L2 ; .....
To see the actual compiler-macro expansion, use a function
(let ((*features* (cons :inline-generic-function *features*))) (print (inline-generic-function '(plus a b)))) ;; Inlining a generic function PLUS (LET ((#:A1734 (1+ A)) (#:B1735 (1- B))) (EMATCH* (#:A1734 #:B1735) (((TYPE FLOAT) (TYPE FLOAT)) (LET ((A #:A1734) (B #:B1735)) (DECLARE (TYPE FLOAT A)) (DECLARE (TYPE FLOAT B)) (+ A B) (LET ((A #:A1734) (B #:B1735)) (DECLARE (TYPE FLOAT A)) (DECLARE (TYPE FLOAT B)) (+ A B)))) (((TYPE FIXNUM) (TYPE FIXNUM)) (LET ((A #:A1734) (B #:B1735)) (DECLARE (TYPE FIXNUM A)) (DECLARE (TYPE FIXNUM B)) (+ A B) (LET ((A #:A1734) (B #:B1735)) (DECLARE (TYPE FIXNUM A)) (DECLARE (TYPE FIXNUM B)) (+ A B))))))
ematch from Trivia pattern matcher expands into thoroughly typed
dispatching code, a sufficiently smart compiler would compile
machine assembly, which is the case at least in SBCL.
Automatic compile-time dispatching
If the code is inlined in a typed environment, smart compilers like sbcl can detect certain branches are not reachable, thus removing the checks and reducing the code size. This is equivalent to compile-time dispatch. In the example below, the code for dispatching FLOAT is removed.
(defun func-using-inlined-plus-and-type-added (a b) " ; disassembly for FUNC-USING-INLINED-PLUS-AND-TYPE-ADDED ; Size: 29 bytes. Origin: #x10031E7788 ; 88: 4801F9 ADD RCX, RDI ; no-arg-parsing entry point ; 8B: 488BD1 MOV RDX, RCX ; 8E: 48D1E2 SHL RDX, 1 ; 91: 710C JNO L0 ; 93: 488BD1 MOV RDX, RCX ; 96: 41BB70060020 MOV R11D, 536872560 ; ALLOC-SIGNED-BIGNUM-IN-RDX ; 9C: 41FFD3 CALL R11 ; 9F: L0: 488BE5 MOV RSP, RBP ; A2: F8 CLC ; A3: 5D POP RBP ; A4: C3 RET " (declare (inline plus)) (declare (optimize (speed 3) (safety 0))) (declare (type fixnum a b)) (plus a b))
If the types does not match, errors are signalled by
which is consistent with the behavior of standard generic functions.
Enabling Inlining Globally
Inlining is not globally enabled by default. This is because the inlined code becomes obsoleted when the generic function definition changes, and therefore you generally do not want to make them inlined during the development.
It can be enabled globally
:inline-generic-function flag in
*features*, which is useful when you build a standalone binary.
When this feature is present, all inlinable generic functions
are inlined unless it is declared
(push :inline-generic-function *features*)
We tested two generic functions, one of which is a standard-generic-function, and another is an inlined-generic-function.
Both generic functions follow the definition below:
(defgeneric plus (a b) [(:generic-function-class inlined-generic-function)]) (defmethod plus :around ((a number) (b number)) (+ a b) (call-next-method)) (defmethod plus ((a fixnum) (b fixnum)) (+ a b)) (defmethod plus ((a double-float) (b double-float)) (+ a b))
We tested them with and without
inline declaration, i.e.,
(defun func-using-plus (a b) (declare (optimize (speed 3) (safety 0))) (plus a b)) (defun func-using-inlined-plus (a b) (declare (inline plus)) (declare (optimize (speed 3) (safety 0))) (plus a b))
Thus, we have 4 configurations in total. The experiment is run under AMD Phenom II X6 processor 2.8GHz with SBCL 1.3.1 (launched by Roswell). The benchmark function is shown below:
(defvar *input* (iter (repeat 1000) (collect (cons (random 100.0d0) (random 100.0d0))) (collect (cons (+ 20 (random 100)) (+ 20 (random 100)))))) (defun benchmark () (time (iter (for (a . b) in *input*) (func-using-normal-plus a b))) (time (iter (for (a . b) in *input*) (func-using-normal-inlined-plus a b))) (time (iter (for (a . b) in *input*) (func-using-plus a b))) (time (iter (for (a . b) in *input*) (func-using-inlined-plus a b))))
We first run the benchmark function 1000 times in order to calibrate the CPU cache. We then run the gc and invoke the benchmark function once more. We use the result of this final run in order to make sure the machine state is stabilized.
Since the difference in the runtime is relatively small due to the small
amount of computation, we consider the processor cycles only. We found
that the cost of generic function invocation is considerably low when an
inlined-generic-function is invoked with
|metaclass and inline declaration||processor cycles||consing|
|standard-generic-function, not inlined||742,285||0|
|inlined-generic-function, not inlined||7,865,080||523,760|
Note that the third case, where the
inlined-generic-function is not
inlined, is slower than the normal generic function. This would be because
we use the non-standard metaclass for representing the generic function and
the normal optimization provided by the implementation is not performed.
However, this is not a problem because we consider the third case only takes
place during the development.
We showed that … well, anyway, this is not a paper. Enjoy!
This library is at least tested on implementation listed below:
- SBCL 1.3.1 on X86-64 Linux 3.19.0-39-generic (author’s environment)
Also, it depends on the following libraries:
- trivia by Masataro Asai
- NON-optimized pattern matcher compatible with OPTIMA, with extensible optimizer interface and clean codebase
- closer-mop by Pascal Costanza
- Closer to MOP is a compatibility layer that rectifies many of the absent or incorrect CLOS MOP features across a broad range of Common Lisp implementations.
- alexandria by
- Alexandria is a collection of portable public domain utilities.
- iterate by
- Jonathan Amsterdam’s iterator/gatherer/accumulator facility
When you use this library as part of your system, make sure that the method definitions are re-evaluated in load-time. This is necessary because the inlining information for the method could be lost after the compilation, i.e., the FASL file does not keep the defmethod form that should be inlined later.
The only thing you need is:
(eval-when (:compile-toplevel :load-toplevel :execute) (defmethod ...) ...)
- Masataro Asai (firstname.lastname@example.org)
Copyright (c) 2015 Masataro Asai (email@example.com)
Licensed under the LLGPL License.