Batch compile_ops across all modules from the root module

[pfultz2]

Motivation

compile_ops ran once per module, so each submodule compiled and benchmarked its own precompile ops in isolation. Identical kernels in different submodules only avoided a second compile by hitting the in-memory binary cache after the first module finished, and the parallel compile was split into many small per-module batches instead of one large one.

Technical Details

• compile_ops::apply now returns immediately for non-root modules. On the root it collects the root plus every reachable submodule, deduplicated and skipping bypassed modules, matching what the pass manager skips.
• compile_manager gained add_plans(ctx, module), and compile() no longer takes a module or is_root flag. Each compile_plan already records its owning module, so results are replaced into cp.mod.
• replace_inserted_device_ops takes a list of modules. It runs dead_code_elimination and lower_device_ops on each, then compiles the inserted device ops in one batch as well. The benchmark MXR dump path passes a single-module list.
• Added same_kernel_in_two_submodules_compiles_once to test/gpu/binary_cache.cpp. An if program with the same pointwise kernel in both branches must produce one code object per submodule with exactly one compile and zero cache reuses, which pins the batched behavior rather than the old compile-then-reuse sequence.

Changelog Category

Add a CHANGELOG.md entry for any option other than Not Applicable

• • Added: New functionality.
• • Changed: Changes to existing functionality.
• • Removed: Functionality or support that has been removed. (Compared to a previous release)
• • Optimized: Component performance that has been optimized or improved.
• • Resolved Issues: Known issues from a previous version that have been resolved.
• • Not Applicable: This PR is not to be included in the changelog.

Follow the LLVM AI Tool Use Policy for contributions using AI.

[Copilot]

🟡 Changes recommended

Nested compilation exits before benchmark MXR files can be saved when dump mode is enabled.

2 open findings

• Prevent nested compilation from exiting before benchmark MXR is saved · New
• Inline the one-off graph builder in the compilation test · New

What changed in this PR

Batches GPU operation compilation across the root module and reachable submodules to improve kernel deduplication and parallelism.

Changes:

• Collects eligible modules and compiles their plans together.
• Batches compilation of inserted device operations.
• Adds cross-submodule binary-cache coverage.

Review was performed as a single pass without agent fan-out.

File	Description
src/​targets/​gpu/​compile_ops.cpp	Implements cross-module compilation batching.
test/​gpu/​binary_cache.cpp	Tests deduplication across conditional submodules.

🧠 Review effort: Balanced

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #5370      +/-   ##
===========================================
- Coverage    92.71%   92.70%   -0.00%     
===========================================
  Files          635      635              
  Lines        36204    36204              
===========================================
- Hits         33563    33562       -1     
- Misses        2641     2642       +1     

see 1 file with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.