Apply BOLT optimizations to libpython for shared builds. Most of the C
code is in libpython so it is critical to apply BOLT there fully realize
BOLT benefits.
This change also reworks how BOLT instrumentation is applied. It
effectively removes the readelf based logic added in gh-101525 and
replaces it with a mechanism that saves a copy of the pre-bolt binary
and restores that copy when necessary. This allows us to perform BOLT
optimizations without having to manually delete the output binary to
force a new bolt run.
Also:
- add a clean-bolt target for purging BOLT files and hook that up to the
clean target
- .gitignore BOLT related files
Before and after this refactor, `make` will no-op after a previous run.
Both versions should also share common make DAG deficiencies where
targets fail to trigger as often as they need to or can trigger
prematurely in certain scenarios. e.g. after this change you may need to
`rm profile-bolt-stamp` to force a BOLT run because there aren't
appropriate non-phony targets for BOLT's make target to depend on.
To make it easier to iterate on custom BOLT settings, the flags to pass
to instrumentation and application are now defined in configure and can
be overridden by passing BOLT_INSTRUMENT_FLAGS and BOLT_APPLY_FLAGS.
This merges their code. They're backed by the same single HACL* static library, having them be a single module simplifies maintenance.
This should unbreak the wasm enscripten builds that currently fail due to linking in --whole-archive mode and the HACL* library appearing twice.
Long unnoticed error fixed: _sha512.SHA384Type was doubly assigned and was actually SHA512Type. Nobody depends on those internal names.
Also rename LIBHACL_ make vars to LIBHACL_SHA2_ in preperation for other future HACL things.
Replace the builtin hashlib implementations of SHA2-384 and SHA2-512
originally from LibTomCrypt with formally verified, side-channel resistant
code from the [HACL*](https://github.com/hacl-star/hacl-star/) project.
The builtins remain a fallback only used when OpenSSL does not provide them.
replacing hashlib primitives (for the non-OpenSSL case) with verified implementations from HACL*. This is the first PR in the series, and focuses specifically on SHA2-256 and SHA2-224.
This PR imports Hacl_Streaming_SHA2 into the Python tree. This is the HACL* implementation of SHA2, which combines a core implementation of SHA2 along with a layer of buffer management that allows updating the digest with any number of bytes. This supersedes the previous implementation in the tree.
@franziskuskiefer was kind enough to benchmark the changes: in addition to being verified (thus providing significant safety and security improvements), this implementation also provides a sizeable performance boost!
```
---------------------------------------------------------------
Benchmark Time CPU Iterations
---------------------------------------------------------------
Sha2_256_Streaming 3163 ns 3160 ns 219353 // this PR
LibTomCrypt_Sha2_256 5057 ns 5056 ns 136234 // library used by Python currently
```
The changes in this PR are as follows:
- import the subset of HACL* that covers SHA2-256/224 into `Modules/_hacl`
- rewire sha256module.c to use the HACL* implementation
Co-authored-by: Gregory P. Smith [Google LLC] <greg@krypto.org>
Co-authored-by: Erlend E. Aasland <erlend.aasland@protonmail.com>
Fix the gdbm_compat library detection logic to actually check for
-lgdbm_compat independently of the ndbm detection.
This fixes the build failure with `--with-dbmliborder=gdbm`,
and implicit fallback to ndbm with the default value.
When cross-compiling, the compile/run test for -pthread always fails so -pthread
will never be automatically set without an override from the cache. ac_cv_pthread
can already be overridden, so do the same thing for ac_cv_cxx_thread.
It has had no effect on non-macOS platforms for a long time, and has had
the non-obvious effect of invoking `pkg_config` and not setting
`-DUSING_APPLE_OS_LIBFFI` on macOS since GH-22855.
This got introduced in commit 5884449539
to determine if readline is already linked against curses or tinfo in
the setup.py, which is no longer present.
Added os.setns and os.unshare to easily switch between namespaces
on Linux.
Co-authored-by: Christian Heimes <christian@python.org>
Co-authored-by: CAM Gerlach <CAM.Gerlach@Gerlach.CAM>
Co-authored-by: Victor Stinner <vstinner@python.org>
⚠️⚠️ Note for reviewers, hackers and fellow systems/low-level/compiler engineers ⚠️⚠️
If you have a lot of experience with this kind of shenanigans and want to improve the **first** version, **please make a PR against my branch** or **reach out by email** or **suggest code changes directly on GitHub**.
If you have any **refinements or optimizations** please, wait until the first version is merged before starting hacking or proposing those so we can keep this PR productive.
- support EMSDK tot-upstream and git releases
- allow WASM assents for wasm64-emscripten and WASI. This makes single file distributions on WASI easier.
- decouple WASM assets from browser builds
* Add support for the BOLT post-link binary optimizer
Using [bolt](https://github.com/llvm/llvm-project/tree/main/bolt)
provides a fairly large speedup without any code or functionality
changes. It provides roughly a 1% speedup on pyperformance, and a
4% improvement on the Pyston web macrobenchmarks.
It is gated behind an `--enable-bolt` configure arg because not all
toolchains and environments are supported. It has been tested on a
Linux x86_64 toolchain, using llvm-bolt built from the LLVM 14.0.6
sources (their binary distribution of this version did not include bolt).
Compared to [a previous attempt](https://github.com/faster-cpython/ideas/issues/224),
this commit uses bolt's preferred "instrumentation" approach, as well as adds some non-PIE
flags which enable much better optimizations from bolt.
The effects of this change are a bit more dependent on CPU microarchitecture
than other changes, since it optimizes i-cache behavior which seems
to be a bit more variable between architectures. The 1%/4% numbers
were collected on an Intel Skylake CPU, and on an AMD Zen 3 CPU I
got a slightly larger speedup (2%/4%), and on a c6i.xlarge EC2 instance
I got a slightly lower speedup (1%/3%).
The low speedup on pyperformance is not entirely unexpected, because
BOLT improves i-cache behavior, and the benchmarks in the pyperformance
suite are small and tend to fit in i-cache.
This change uses the existing pgo profiling task (`python -m test --pgo`),
though I was able to measure about a 1% macrobenchmark improvement by
using the macrobenchmarks as the training task. I personally think that
both the PGO and BOLT tasks should be updated to use macrobenchmarks,
but for the sake of splitting up the work this PR uses the existing pgo task.
* Simplify the build flags
* Add a NEWS entry
* Update Makefile.pre.in
Co-authored-by: Dong-hee Na <donghee.na92@gmail.com>
* Update configure.ac
Co-authored-by: Dong-hee Na <donghee.na92@gmail.com>
* Add myself to ACKS
* Add docs
* Other review comments
* fix tab/space issue
* Make it more clear that --enable-bolt is experimental
* Add link to bolt's github page
Co-authored-by: Dong-hee Na <donghee.na92@gmail.com>
Victor Stinner
Dong-hee Na
T. Wouters
Gregory Szorc
Erlend E. Aasland
Zhang Na
Petr Viktorin
Nikita Sobolev
Matthias Görgens
Jonathan Protzenko
Eli Schwartz
Gregory P. Smith
Eric Snow
Kushal Das
Michał Górny
Dustin Spicuzza
Zachary Ware
Sam James
colorfulappl
Miro Hrončok
Noam Cohen
Sam James
serge-sans-paille
Jelle Zijlstra
Pablo Galindo Salgado
Christian Heimes
Kevin Modzelewski