logseq/pages/hls__sle23-paper95_1682888032661_0.md

file:: sle23-paper95_1682888032661_0.pdf file-path:: ../assets/sle23-paper95_1682888032661_0.pdf

• approach to define an executable semantics targeting the development of optimizing compilers. ls-type:: annotation hl-page:: 1 hl-color:: green id:: 644ed678-8c79-42a6-a523-2cea2d8ef419
• ype checks ls-type:: annotation hl-page:: 1 hl-color:: green id:: 644ed708-429c-49c9-95d3-39140d8b08f2
• primitive ls-type:: annotation hl-page:: 1 hl-color:: green id:: 644ed70b-5f81-4cf4-998e-8b60e9879bf1
• alues boxing and unboxing ls-type:: annotation hl-page:: 1 hl-color:: green id:: 644ed70f-1d40-476a-8baf-2c841cf958ab
• function calls ls-type:: annotation hl-page:: 1 hl-color:: green id:: 644ed716-2eb6-4108-8d99-95474df37342
• semPy, a partial evaluator of our executable semantics ls-type:: annotation hl-page:: 1 hl-color:: purple id:: 644ed738-190c-4729-b578-00c935ec543c
• On some tasks, Zipi displays performance competitive with that of state of art Python implementations ls-type:: annotation hl-page:: 1 hl-color:: purple id:: 644ed74b-53ad-44c3-ae5e-2f1801a6b0d3
• While the syntax is formally specified, this is not the case of the semantics, leaving room for ambiguities and making it difficult to reason about Python ls-type:: annotation hl-page:: 1 hl-color:: purple id:: 644ed772-c110-49c2-af3b-7cc5df1a8792
• CPython, the reference implementation ls-type:: annotation hl-page:: 1 hl-color:: purple id:: 644ed797-342f-4736-9ef3-7c4e3b5db4b6
• RPython experiment ls-type:: annotation hl-page:: 1 hl-color:: green id:: 644ed7b1-91e2-4842-b8d9-938da5e84915
• development of an executable semantics, written in the Python syntax, that aims to automate part of the implementation of a Python optimizing compiler. ls-type:: annotation hl-page:: 1 hl-color:: purple id:: 644ed7c1-2f39-45db-a99c-a11a463fb4ad
• overview of Python’s semantics ls-type:: annotation hl-page:: 1 hl-color:: green id:: 644ed7e1-a4e8-4625-ba5e-e78640abda1d
• executable semantics that describes the behavior of various Python operations. ls-type:: annotation hl-page:: 1 hl-color:: green id:: 644ed7e8-7ea9-4d0f-b32c-8443ce2bfd91
• technique for partial evaluation of our executable semantics that focuses on removing redundant type checks, boxing and unboxing, and method lookups and invocations from Python operations. ls-type:: annotation hl-page:: 1 hl-color:: green id:: 644ed7f3-c2b3-4d30-8662-a0250b933bb8
• we show how we reused our executable semantics in the implementation of Zipi ls-type:: annotation hl-page:: 1 hl-color:: green id:: 644ed82f-68ac-4a06-8ea0-4e18cffa3bf6
• Python’s semantics is highly dynamic ls-type:: annotation hl-page:: 1 hl-color:: purple id:: 644ed8b2-7141-4433-9de8-8d2fed9350a6
• Python does not have an official formal semantics. ls-type:: annotation hl-page:: 1 hl-color:: purple id:: 644ed8bd-bf6e-43e7-9224-6441ca9d6a37
• we refer to the behavior of CPython ls-type:: annotation hl-page:: 1 hl-color:: purple id:: 644ed8ce-3fff-49a1-8e5c-d211c1596856
• identity ls-type:: annotation hl-page:: 2 hl-color:: green id:: 644edbe1-425d-4f9f-953f-70a1eee826c5
• value ls-type:: annotation hl-page:: 2 hl-color:: green id:: 644edbe4-8fcb-46c2-8f5f-d6b37e62d9bf
• type determines the operations allowed on the object. ls-type:: annotation hl-page:: 2 hl-color:: green id:: 644edc32-9694-427c-b824-203f2500c993
• However, this is not possible for objects whose type is a built-in type such as booleans, floats, integers, strings, lists, tuples, sets and dictionaries ls-type:: annotation hl-page:: 2 hl-color:: green id:: 644edc47-bfd5-47d1-9615-9d964a0fc4eb
• All values in a Python program are objects. ls-type:: annotation hl-page:: 2 hl-color:: green id:: 644edc50-0c11-42fa-a519-2c0a86cb13c7
• The expression type(x).add first looks for add on type(x) itself. If no such method is found, it is looked up recursively on the parents of type(x) ls-type:: annotation hl-page:: 2 hl-color:: green id:: 644fdde3-026c-4d9e-8491-9287b192d97e
• method resolution order (MRO ls-type:: annotation hl-page:: 2 hl-color:: green id:: 644fddef-928d-4d60-889a-275563a5e3b0
• The values of their attributes may be updated, new attributes may be introduced, or existing ones may be removed ls-type:: annotation hl-page:: 2 hl-color:: green id:: 644fde76-2be1-4d63-8c9f-6b38110e595d
• An important exception is that attributes of all built-in types are read-only ls-type:: annotation hl-page:: 2 hl-color:: green id:: 644fde84-dc3e-4011-bcb4-53ed1ec50971
• Immutability of built-in types is part of Python’s semantics. ls-type:: annotation hl-page:: 2 hl-color:: green id:: 644fde97-7df2-4cd3-be39-61207bf1012b
• Python incorporates features such as dynamic typing, dynamic binding, and dynamic code evaluation. ls-type:: annotation hl-page:: 2 hl-color:: green id:: 644fdfa2-cf9f-4cf7-8a6c-3c18dfcfca98
• Python supports modular programming through module objects ls-type:: annotation hl-page:: 3 hl-color:: green id:: 644fe2a1-785f-467f-a592-d08d056273f8
• any modifications applied to a module’s attributes is reflected on its global environment. ls-type:: annotation hl-page:: 3 hl-color:: green id:: 644fe2b7-09a9-4c4d-9acd-b237e76f9e6e
• An operation as simple as subtracting an integer and a floating point number requires two method searches in the MRO of int and float respectively ls-type:: annotation hl-page:: 3 hl-color:: green id:: 644fe32c-2c30-4f7b-a62b-6e740481f75a
• This procedure, known as boxing and unboxing, leads to additional overhead [5] ls-type:: annotation hl-page:: 3 hl-color:: green id:: 644fe360-695a-4a9b-8f4f-44a1e790e943
• xecutable semantics aimed at developing optimizing Python compilers ls-type:: annotation hl-page:: 3 hl-color:: green id:: 644fe36e-583b-4982-b50a-3efa431d3d0b
• 1. automate the implementation of a PYTHON compiler, ls-type:: annotation hl-page:: 3 hl-color:: yellow id:: 644fe378-f02f-4626-bb6b-1b797d6c12b3 hl-stamp:: 1682957179733
• be easily reusable by existing PYTHON compilers ls-type:: annotation hl-page:: 3 hl-color:: green id:: 644fe3a1-93fb-427e-aec2-147425967c89 hl-stamp:: 1682957221312
• yield performant implementations ls-type:: annotation hl-page:: 3 hl-color:: yellow id:: 644fe3ad-606d-4a0f-b985-8134947edfcd hl-stamp:: 1682957231924
• RPython ls-type:: annotation hl-page:: 3 hl-color:: yellow id:: 644fe5dc-a608-41ec-90ac-479206bfdf0c
• It differs in that we instead ls-type:: annotation hl-page:: 3 hl-color:: yellow id:: 644fe60f-0f34-42c5-a2bc-2cb6ac4060f0
• compiler intrinsics statement ls-type:: annotation hl-page:: 4 hl-color:: yellow id:: 644fe635-ccc4-4ded-9720-af2852297010
• Intrinsics imported with the compiler intrinsics statement are static, they cannot be shadowed by another assignment or assigned to a variab ls-type:: annotation hl-page:: 4 hl-color:: green id:: 644fe6e2-b255-472c-ae84-a88032fe6835 hl-stamp:: 1682958056044
• addition ls-type:: annotation hl-page:: 4 hl-color:: green id:: 644fe729-e3a8-4f0b-a916-b478f3a6370f
• These examples effectively illustrate why seemingly simple operations incur a significant overhead. ls-type:: annotation hl-page:: 4 hl-color:: green id:: 644fe73b-7ce7-4159-8c24-735bba4f43be
• [:span] ls-type:: annotation hl-page:: 2 hl-color:: green id:: 644feb0a-3015-4266-907b-40e6abd43436 hl-type:: area hl-stamp:: 1682959114081
• class_getattr ls-type:: annotation hl-page:: 4 hl-color:: green id:: 6450131f-fb3a-445f-a406-c06df5147bea
• absent ls-type:: annotation hl-page:: 4 hl-color:: green id:: 64501321-fece-4916-90c2-f9dd0f7cb56c
• normal ls-type:: annotation hl-page:: 4 hl-color:: green id:: 64501669-7f6e-4d90-852a-4f0f2b0f3cc0
• reflected ls-type:: annotation hl-page:: 4 hl-color:: green id:: 6450166b-efb7-48b3-9fcd-91e1c789d9b1
• a semantics ls-type:: annotation hl-page:: 4 hl-color:: green id:: 64501688-27cb-46fb-8e64-bb7d280a6b77
• semantics of ls-type:: annotation hl-page:: 4 hl-color:: green id:: 6450168b-d80b-4701-a50a-b0981968b34b
• This allows us to define the notion of behavior, a specialization of an operator for a given combination of built-in types. ls-type:: annotation hl-page:: 12 hl-color:: green id:: 64501799-4f47-4d73-af89-3a736156bf15
• Zipi compiles behaviors and dispatches operations to their corresponding behaviors at run time. ls-type:: annotation hl-page:: 12 hl-color:: purple id:: 645017bf-849d-4665-a879-91a3aa34ab76
• PyPy ls-type:: annotation hl-page:: 12 hl-color:: yellow id:: 645017c6-757f-40a8-8461-4f7fb63479dd
• It appears to us that they would be well suited for CPython, as they specifically address the known overhead of this implementation ls-type:: annotation hl-page:: 12 hl-color:: green id:: 645017ec-07e4-4eb9-a940-cb54f7c71008
• This section presents semPy3, a Python tool for generating behaviors by removing redundant type checks, boxing and unboxing, and method calls whenever possible ls-type:: annotation hl-page:: 7 hl-color:: green id:: 64501b49-37e7-43f6-a9a8-6cd9548c5831
• semPy is a Python partial evaluator supporting the compiler intrinsics statement. ls-type:: annotation hl-page:: 7 hl-color:: yellow id:: 64501b57-a394-4c43-9dd4-f302dd17c5d3
• It outputs a specialization of the semantics given that context, which is a behavior. ls-type:: annotation hl-page:: 7 hl-color:: purple id:: 64501b77-d4f6-400a-869d-8c88ca84fe5b
• add_intX_floatY was generated from the add semantics (Figure 3) in a context where the left-hand operand is an int and the right-hand operand is a float ls-type:: annotation hl-page:: 7 hl-color:: blue id:: 64501b97-fc3f-4ea7-ac34-bec5ce768cb9 hl-stamp:: 1682971545826
• we configured our benchmarks to only measure the run time performance of the program after initialization ls-type:: annotation hl-page:: 11 hl-color:: green id:: 6451854d-ddd6-4638-b46a-065d44249997
• or ls-type:: annotation hl-page:: 11 hl-color:: red id:: 6451856b-a483-4390-8b4a-5a1513822cd2
• operations suffering from poor performance. ls-type:: annotation hl-page:: 11 hl-color:: green id:: 645185b1-4388-4327-b89f-dc3b7c9c9f50
• The operation being evaluated is wrapped in a loop to reach a measurable time on the order of one second on CPython. ls-type:: annotation hl-page:: 11 hl-color:: green id:: 645185be-c478-4062-b66d-62a1b30f8623
• uncompetitive in comparison to static languages such as C ls-type:: annotation hl-page:: 1 hl-color:: green id:: 6452262b-c2a5-45e9-936e-b1c6aafb4a04
• The effort spent getting the semantics right leaves little time for optimization ls-type:: annotation hl-page:: 1 hl-color:: blue id:: 645228f1-38c2-4ae0-b579-823a46cf2b28 hl-stamp:: 1683106035825
• methods governing operations are called magic methods. ls-type:: annotation hl-page:: 2 hl-color:: purple id:: 64522a40-df2b-40c5-ac97-11a12e2612b1 hl-stamp:: 1683106371076
• The only operators that cannot be overloaded are the “is” operator which compares objects by identity, and the “and”,“or” and “not” boolean operators ls-type:: annotation hl-page:: 2 hl-color:: purple id:: 64522aaa-e11f-43f2-b5de-ee3550d58fea
• automate this process to accelerate development, including that of existing compilers, independently of the language and tools chosen for its implementation. ls-type:: annotation hl-page:: 3 hl-color:: purple id:: 64522e0a-730b-4302-97c2-cb1bd4b3acca
• interface with the semantics by using the parsing infrastructure of an existing compiler. ls-type:: annotation hl-page:: 3 hl-color:: yellow id:: 64522e83-2ecd-4c6a-bacf-e604cfeb447d hl-stamp:: 1683107462697
• optimized versions of various operators in sections 5 and ls-type:: annotation hl-page:: 4 hl-color:: green id:: 64522ec4-6715-4632-9a37-71aef6cd8a67 hl-stamp:: 1683107527781
• Behaviors ls-type:: annotation hl-page:: 6 hl-color:: green id:: 645232f1-27ff-4c60-a8aa-e9ad76f7eff9
• A compiler can implement arithmetic operators from the semantics defined in Section 3. Yet, by doing so in a naive way, that is calling each magic method, the implementation would likely offer poor performance. ls-type:: annotation hl-page:: 6 hl-color:: green id:: 6452331a-4d1a-48c4-b3d9-7c0e01558c82
• We exploit that fact to generate optimized versions of Python operators. ls-type:: annotation hl-page:: 6 hl-color:: green id:: 64523401-621a-4d9c-af49-0bbc838d177b
• When a semantics or magic method is invoked, semPy systematically inlines the callee’s code at the CALL site. This removes magic method calls from semantics specializations. Magic methods are returned by invocations of the class_getattr intrinsic function. This function is always called on the arguments of a semantics, whose types are provided in the type context, so it is always possible to resolve which magic method is to be called, or if that method is absent. ls-type:: annotation hl-page:: 7 hl-color:: yellow id:: 645270f9-9991-40aa-b8cf-42d909fc85fd hl-stamp:: 1683124476488
• optimizing PYTHON compiler. ls-type:: annotation hl-page:: 9 hl-color:: purple id:: 64527140-4f7d-40d0-a58c-4db765477243 hl-stamp:: 1683124552273
• Zipi compiles PYTHON to Scheme code, which is then compiled to an executable using either the Bigloo [7] or Gambit [8] Scheme compilers ls-type:: annotation hl-page:: 9 hl-color:: yellow id:: 64527182-5134-4f53-9126-eb3069193e4f hl-stamp:: 1683124613750
• Performance was measured through microbenchmarks as well as regular benchmarks implementing well-known algorithms ls-type:: annotation hl-page:: 11 hl-color:: yellow id:: 6452738f-a8b3-42c8-bbe8-6887ded91ba9
• Zipi has a significant compilation overhead because it is AOT and has a deep pipeline that compiles Python code to Scheme, then to C, and finally to machine code. ls-type:: annotation hl-page:: 11 hl-color:: yellow id:: 645273a0-f0c0-4759-ae0d-d9717fe0babb
• Benchmarks measure real time using the Python time module which all implementations provide. ls-type:: annotation hl-page:: 11 hl-color:: yellow id:: 6452ba2e-27ba-45bc-bb08-c97e83fa8d55
• To minimize the loop overhead, its body contains several repetitions of the measured operation (typically 20 ls-type:: annotation hl-page:: 11 hl-color:: yellow id:: 6452bd0c-bde8-477e-b00a-7e3b2747b09f
• Unfortunately, it does not allow a comparison with PyPy which treats the kernel of many of our benchmarks as dead cod ls-type:: annotation hl-page:: 11 hl-color:: yellow id:: 6452bd82-bae0-4c14-b0e2-c8a13a6dd9db
• Performance improvements from other optimizations unrelated to behaviors also show up in the microbenchmarks ls-type:: annotation hl-page:: 11 hl-color:: yellow id:: 6452bec7-50a0-4330-b516-8e8d2dd905db
• For instance, assignment to global variables, function calls and iteration on built-in types are all faster than with CPython ls-type:: annotation hl-page:: 11 hl-color:: yellow id:: 6452becd-3128-4dd1-a6df-c0913f0e70cc
• On the other hand, some microbenchmarks display poor performance. Those are unoptimized features that we implemented in a naive way, such as function calls with keyword arguments. ls-type:: annotation hl-page:: 11 hl-color:: yellow id:: 6452bed2-863a-4a94-becb-c70fe197dbb3
• Zipi being at an early development stage, only four benchmarks from PyPerformance are supported at the moment, hence the need for custom benchmarks ls-type:: annotation hl-page:: 11 hl-color:: yellow id:: 6452beee-2cf9-4990-b595-620cd469b4e5
• We wish to extend the behavior optimization to other operations in the future to further analyze its impact on performance ls-type:: annotation hl-page:: 12 hl-color:: yellow id:: 6452bf9c-18e8-4d21-99ea-58199131bde6