gh-139109: A new tracing JIT compiler frontend for CPython

[Fidget-Spinner]

This PR changes the current JIT model from trace projection to trace recording. Benchmarking: 0.6% slower pyperformance versus TC+current JIT, However, 70% faster Richards on the most improved benchmark versus the current JIT. https://github.com/facebookexperimental/free-threading-benchmarking/blob/main/results/bm-20251020-3.15.0a1%2B-ec2971f-CLANG%2CJIT/bm-20251020-vultr-x86_64-Fidget%252dSpinner-tracing_jit-3.15.0a1%2B-ec2971f-vs-base.md

This new JIT frontend is already able to record/execute significantly more instructions than the previous JIT frontend. In this PR, we are now able to record through custom dunders, simple object creation, generators, etc. None of these were done by the old JIT frontend. Some custom dunders uops were discovered to be broken as part of this work gh-140277

Some tests for the optimizers are disabled for now not because the optimizer doesn't work, but that it's actually unrolling recursion, thus leading to different results in the trace!

The full test suite passes on my system.

Not close enough to a holiday, so no poem this time!

Pros:

• Ignoring the generated tracer code as it's automatically created, this is only additional 1k lines of code. The maintenance burden is handled by the DSL and code generator.
• optimizer.c is now significantly simpler, as we don't have to do strange things to recover the bytecode from a trace.
• The new JIT frontend is able to handle a lot more control-flow than the old one.
• Tracing is very low overhead. We use the tail calling interpreter/computed goto interpreter to switch between tracing mode and non-tracing mode. Specialization is still enabled while tracing.
• Better handling of polymorphism. We leverage the specializing interpreter for this.

Cons:

• (For now) requires tail calling interpreter or computed gotos. This means no Windows JIT for now :(. Not to fret, tail calling is coming soon to Windows though gh-139922: Tail calling for MSVC (VS 2026) #139962
• Build times for JIT is now a lot higher as we build two interpreters, not one. computed goto interpreter's build times is very long on my system. Tail call is roughly half the build times of computed goto as there's no blowup of CFG. I recommend building with tail call in the future if you want fast build times.

Design:

• We (ab)use the cases generator DSL to duplicate the normal interpreter --- one specializing interpreter, one tracing version.
• The tracing version behaves nearly identical to the normal interpreter, in fact it even has specialization! This allows it to run without much of a slowdown when tracing. The actual cost of tracing is only a function call and writes to memory.
• This was used instead of not duplicating the interpreter as it offers more powerful customization and specialization (see below).
• The tracing interpreter uses the specializing interpreter's deopt to naturally form the side exit chains. This allows it to side exit chain effectively, without repeating much code. We force a re-specializing when tracing a deopt.
• The tracing interpreter can even handle goto errors/exceptions, but I chose to disable them for now as it's not tested. This is the power of having another interpreter.
• Because we do not share interpreters, there is no measurable slowdown to the original specializing interpreter (on Tailcall only, the design needs to be changed for computed goto). With JIT disabled. With JIT enabled, there might be a slowdown in the form of the JIT trying to trace.
• Things that could have dynamic instruction pointer effects are guarded on. The guard deopts to a new instruction --- _DYNAMIC_EXIT. This new instruction will cause a re-trace when hot enough starting from the target instruction.
• As such, traces "grow" like a web/graph rather than a tree. It starts from a loop, then grows outwards naturally consuming more code around it as the hotspot gets hotter.
• I believe this might be the first time trace recording is combined with an inline caching + quickening interpreter (in literature by Brunthaler at least it's called that) to achieve specialization. Let's hope this goes well.

• Issue: Trace recording JIT #139109

[Fidget-Spinner]

Closing because this is triggering some UB in CI and wasting resources. Will reopen once I fix it.

[markshannon]

This is exciting stuff.
I've done a preliminary review. More to come later.

@Fidget-Spinner do you have any ideas as to why the go benchmark is 20% slower?

[markshannon]

This is correct, but it made me curious.
OOI, did you just spot this, or did the JIT manage to end up running code from a deallocated code object?

[Fidget-Spinner]

The main problem isn't running deallocated code. The problem was that the JIT actually traced then tried to compile code that happened to deallocate while tracing. Thus leading to segfaults while optimizing.

This came up in a few tests from the test suite.

[markshannon]

We may still want to prune the trace in the optimizer. For example, if the optimizer's confidence drops too low.
So leave optimize_uops returning the length.

[Fidget-Spinner]

The current code is wrong. The optimizer never errors. So length can never be 0.

The optimizer currently just returns the length of the unoptimized trace if it doesn't optimize. So this entire code is dead code.

However, I'll leave it there in case you need it in the future.

[markshannon]

We've already got too many macros here, and it makes the code really hard to follow.
It also creates code bloat as we replicate all this code 256 times.

At the very least, factor out as much as you can into helper functions.
Almost everything in this macro after LEAVE_TRACING() could go in a helper function.

[markshannon]

Never more true 🙂

[markshannon]

SHOULD_CONTINUE_TRACING doesn't seem to be used anywhere

[Fidget-Spinner]

It's used by _COLD_EXIT to tell the interpreter to resume tracing upon exiting to tier 1. This is what allows it to form the side exit trace.

[markshannon]

Can we fix up this table, so we don't need all the special cases afterwards?

[Fidget-Spinner]

@markshannon I suspect the greatest slowdown is due to the code being rather branchy. Our side exits are very costly right now still. I am planning to fix this thru #140434.

Other sources of slowdown: we even trace through CALL_ALLOC_AND_ENTER_INIT (simple object initialization and creation) right now. The optimizer can't deal with that, so it stops optimizing there. I have another branch that adds this to the uop optimizer.

Finally, note that I found that the current design slows down the base computed goto interpreter by a lot!!! (Roughly 6% pyperformance it seems). I suspect it's because we trace when we PGO, so we have no choice but to do a normal call to another interpreter when do computed goto rather than the dispatch table idea.

Alternatively, just mandate tail calling for JIT builds.