Author and test XIR transformation passes for the LuisaCompute compiler IR.
npx skills add https://github.com/LuisaGroup/LuisaCompute --skill xir-passes
Please help me install this Agent Skill. Skill: xir_passes Source: https://github.com/LuisaGroup/LuisaCompute/tree/main/.agents/skills/xir_passes Command: npx skills add https://github.com/LuisaGroup/LuisaCompute --skill xir-passes
Writing compiler transformation passes for LuisaCompute's XIR intermediate representation requires deep knowledge of non-obvious API conventions, terminator semantics, and mutation patterns; this Skill captures that hard-won knowledge so you avoid common pitfalls like invalid casts, wrong setter names, and unsafe instruction-list mutation.
<Name>Info POD interface, module/function entry points, PassReport statistics, and PassPipeline integration with fixed-point sub-pipelines.get_memory_info(), plus LLVM pass equivalents for cross-referencing implementations.src/xir/passes/. This Skill tells you exactly where to put the header and implementation, how to register it in CMake, how to write Boost.UT tests with reachable test blocks, and which API traps to avoid.Use the xir_passes skill to help me implement a new XIR transformation pass under src/xir/passes/ with proper registration and unit tests.
High-intent search queries and answers about installing and using this skill.
Create a header in include/luisa/xir/passes/ and implementation in src/xir/passes/, register it in src/xir/CMakeLists.txt, and expose a <Name>Info POD with function-level and module-level entry points. Add tests in src/tests/unit/xir/ following existing test_xir_pass_* files.
Use the two-phase collect-rewrite idiom: first traverse blocks to collect target instructions into a vector, then mutate them in a second loop. Never modify the instruction list during traversal, and wrap in a fixed-point loop if rewrites create new candidates.
No, XIR does not provide cast_or_null or LLVM-style cast. Use v->isa<SomeType>() followed by static_cast<SomeType*>(v), and use func->definition() to safely obtain a FunctionDefinition pointer that returns nullptr for external functions.
traverse_basic_blocks only visits blocks reachable from body_block. Orphan test blocks are silently skipped, so wire them up with br or cond_br from the entry block, for example using a constant-true conditional branch.
Only pure or memory-reading instructions with empty use lists are removable, such as ARITHMETIC, CAST, GEP, LOAD, and RESOURCE_QUERY. Never remove stores, atomics, calls to definitions, prints, or asserts; use get_memory_info() from helpers.h to check effects.
The skill includes a mapping table: dce maps to LLVM DCE.cpp, gvn to GVN.cpp, mem2reg to Mem2Reg.cpp, restructure_cfg to StructurizeCFG.cpp, and licm to LICM.cpp. XIR-specific passes like autodiff and convergence_region have no LLVM equivalent.
awfixers-stuffGuide authoring, building, and debugging C++20 modules across compilers and CMake.
arbisoftAutomate end-to-end C++ testing workflows with GoogleTest and CTest.
backengineeringAnalyze LLVM IR and apply optimization passes to improve generated code.
NeverSightDevelop LLVM-based compiler pipelines from frontends to optimized code generation.