Part 1: Restructuring of compiler pipelines. #3423

schweitzpgi · 2025-09-08T16:22:14Z

There are a variety of compiler pass pipelines that exist in the project. It has long been acknowledged that they don't meet our needs very well.

For the C++ AOT compiler, there are two sets of pipelines primarily. There is the pipeline that is built in the compiler driver and then a pipeline that is used to generate code for default simulation.

For the C++ JIT compiler, every target can (and did) specify its own special pipeline, leading to inconsistencies and maintenance problems. Furthermore, this pipeline may have been incorrect and patched on-the-fly by the JIT compilation process. Finally, it was augmented by running the same pipeline as the AOT compiler for code generation with various minor changes. This situation was both far too flexible (anyone could run any pass in any order in the JIT pipeline) and far too inflexible (high-level and low-level passes were thrown together in ad hoc ways, preventing downstream passes from running correctly).

For Python, the pipelines are done in a completely different way for the most part, but some of the pipelines are the same. For example, the code generation pipeline remains largely the same, but has its own minor tweaks. Python doesn't really distinguish between AOT and JIT compilation.

The changes:

Instead of a single "lowering pipeline" that was used by the JIT compiler, the JIT pipeline has been decomposed into target agnostic and target dependent passes. Furthermore, the JIT pipeline has been broken into 3 levels: high level, mid level, and low level. Each of these levels runs the target dependent passes (from the .yml target config) followed by the target independent passes for that level.
The code generation pipeline now has distinct AOT vs. JIT separation. This eliminates the running of passes at JIT time that are not required.

There are a variety of compiler pass pipelines that exist in the project. It has long been acknowledged that they don't meet our needs very well. For the C++ AOT compiler, there are two sets of pipelines primarily. There is the pipeline that is built in the compiler driver and then a pipeline that is used to generate code for default simulation. For the C++ JIT compiler, every target can (and did) specify its own special pipeline, leading to inconsistencies and maintenance problems. Furthermore, this pipeline may have been incorrect and patched on-the-fly by the JIT compilation process. Finally, it was augmented by running the same pipeline as the AOT compiler for code generation with various minor changes. This situation was both far too flexible (anyone could run any pass in any order in the JIT pipeline) and far too inflexible (high-level and low-level passes were thrown together in ad hoc ways, preventing downstream passes from running correctly). For Python, the pipelines are done in a completely different way for the most part, but some of the pipelines are the same. For example, the code generation pipeline remains largely the same, but has its own minor tweaks. Python doesn't really distinguish between AOT and JIT compilation. The changes: - Instead of a single "lowering pipeline" that was used by the JIT compiler, the JIT pipeline has been decomposed into target agnostic and target dependent passes. Furthermore, the JIT pipeline has been broken into 3 levels: high level, mid level, and low level. Each of these levels runs the target dependent passes (from the .yml target config) followed by the target independent passes for that level. - The code generation pipeline now has distinct AOT vs. JIT separation. This eliminates the running of passes at JIT time that are not required. - Python's compilation model is changed to better reflect that of C++. Compilation is done in 2 phases: AOT and JIT. These phases are not completely custom, but better mirror those of the C++ compilation and execution. (Todo) Signed-off-by: Eric Schweitz <[email protected]>

Signed-off-by: Eric Schweitz <[email protected]>

targettests/TargetConfig/RegressionValidation/anyon.config

runtime/cudaq/platform/default/rest/helpers/ionq/ionq.yml

lib/Optimizer/CodeGen/ConvertToQIRAPI.cpp

github-actions · 2025-09-09T00:52:17Z