Look into string sharing for deep-frozen code

[gvanrossum]

When code objects are created by the compiler or by marshal, certain strings are interned using PyUnicode_InternInPlace(). But code objects obtained from the deep-freeze process do not do this.

While strings are deduped within a module, certain strings (e.g. dunders, 'name', 'self') occur frequently in different modules and thus the deep-frozen code object will use up slightly more space, and certain operations will be slightly slower (e.g. comparison of two strings that are both interned is done by pointer comparison).

We could do a number of different things (or combine several):

• Add PyUnicode_InternInPlace() calls to the "get toplevel" function in each deepfrozen .c file. This would still waste the space though.
• Merge all deepfrozen files into a single file and dedupe strings when that file is written. Requires changing the deepfreeze build procedures for Win and Unix. An advantage is that we could dedupe other things (basically all constants, even bytecode) this way. (@markshannon seems to lean towards this one.)
• Give strings that are likely candidates (i.e., that look like ASCII identifiers -- see all_name_chars() in codeobject.c) an external name with "weak linkage" so that the linker can dedupe them. (Props to @lpereira for this one.)
• For strings that occur in the array of known _Py_Identifiers, replace the string with a reference into that array, for pure savings. (@ericsnowcurrently has more details; IIRC this array isn't in "main" yet.)

[I am not planning to attack this any time soon, so if somebody wants to tackle this, go ahead and assign to yourself.]

[oraluben]

FYI: In our mmap mechanism, we obverse ~10% slowdown in some complex benchmarks, when short path of dict lookup failed because of copied strings. We're in the middle of adding PyUnicode_InternInPlace calls when loading the archive.

Update: some numbers generated by this simple test case https://github.com/oraluben/pyperformance/blob/cd7050101e0c32c0e3fa0359785a4cc28530e3ed/pyperformance/data-files/benchmarks/bm_module_load_attr/run_benchmark.py

*-raw.json are referring clean cpython and *-cds.json are our mmap result.

before:

$pyperf compare_to pyperformance-3.9-raw.json pyperformance-3.9-cds.json pyperformance-3.10-raw.json pyperformance-3.10-cds.json --table
+------------------+-----------------------+-----------------------+------------------------+------------------------+
| Benchmark        | pyperformance-3.9-raw | pyperformance-3.9-cds | pyperformance-3.10-raw | pyperformance-3.10-cds |
+==================+=======================+=======================+========================+========================+
| module_load_attr | 163 ms                | 183 ms: 1.12x slower  | 149 ms: 1.10x faster   | 181 ms: 1.11x slower   |
+------------------+-----------------------+-----------------------+------------------------+------------------------+

after (strings in mapped memory are re-interned):

$pyperf compare_to pyperformance-3.9-raw.json pyperformance-3.9-cds.json pyperformance-3.10-raw.json pyperformance-3.10-cds.json --table
+------------------+-----------------------+-----------------------+------------------------+------------------------+
| Benchmark        | pyperformance-3.9-raw | pyperformance-3.9-cds | pyperformance-3.10-raw | pyperformance-3.10-cds |
+==================+=======================+=======================+========================+========================+
| module_load_attr | 166 ms                | not significant       | 148 ms: 1.12x faster   | 150 ms: 1.11x faster   |
+------------------+-----------------------+-----------------------+------------------------+------------------------+

[kumaraditya303]

Merge all deepfrozen files into a single file and dedupe strings when that file is written. Requires changing the deepfreeze build procedures for Win and Unix. An advantage is that we could dedupe other things (basically all constants, even bytecode) this way. (@markshannon seems to lean towards this one.)

I liked @markshannon idea and I got it working on linux, it reduced the size of generated files by around 1.2Mb as it shares the constants etc. Here is the branch (https://github.com/kumaraditya303/cpython/tree/deepfreeze), next I'll see how to get it working on Windows. (EDIT: It should be Mb not MB)

[kumaraditya303]

Status Update: @gvanrossum I have got it working on Linux and Windows so should I create a PR ?

[pxeger]

@oraluben the line numbers in that link will break when dictobject.c changes. Here's a permalink (once you've highlighted the code range, click the three dots to the left of the line numbers and Copy Permalink):

https://github.com/python/cpython/blob/be578e0c063dad1dbb273f86d5bc77e4e6f14583/Objects/dictobject.c#L794-L795

[gvanrossum]

Status Update: @gvanrossum I have got it working on Linux and Windows so should I create a PR ?

Sounds great! Let's make it a draft PR so I can try it out.

[kumaraditya303]

@gvanrossum Created PR here python/cpython#30572

[gvanrossum]

I see you already submitted a PR and created an issue (https://bugs.python.org/issue46430) to intern strings.

We need another canonicalization: "small integers" should use their cached equivalent.

[kumaraditya303]

We need another canonicalization: "small integers" should use their cached equivalent.

That was much simpler See python/cpython#30715

[kumaraditya303]

Since most of what is mentioned in the issue is done, can the issue be changed with a checklist of all the items and be marked with what is already done ?

[gvanrossum]

Why don't you summarize what you think still needs to be done.

[kumaraditya303]

Current Progress:

• DONE:

  • bpo-46429: Merge all deepfrozen files into one python/cpython#30572

  • bpo-46443: deepfreeze use preallocated small ints python/cpython#30715

  • bpo-46471: use global singletons for single byte bytes objects python/cpython#30781

  • bpo-46608: exclude marshalled-frozen data if deep-freezing to save 300 KB space  python/cpython#31074

  • bpo-46430: Intern strings in deep-frozen modules  python/cpython#30683

  • bpo-46541: Replace core use of _Py_IDENTIFIER() with statically initialized global objects. python/cpython#30928

  • bpo-46753: Add the empty tuple to the _PyRuntimeState.global_objects. python/cpython#31345