llvm-project

795b012e - [mlir][bufferization] Avoid quadratic alias checks for known allocation roots (#242131)

[mlir][bufferization] Avoid quadratic alias checks for known allocation roots (#242131) ### Summary `SplitDeallocWhenNotAliasingAnyOther` currently rebuilds the remaining operand list and performs pairwise alias checks for every memref in a `bufferization.dealloc`. For a wide dealloc with distinct allocation roots, this work scales quadratically with the number of operands. This patch introduces an allocation-root fast path: - Trace each candidate operand through `getViewBase` to a `memref.alloc` root. - Count occurrences of each root using `llvm::DenseMap<Value, int64_t>`. - When all roots are known, determine whether a memref can split off by checking if its root occurs exactly once. - Keep the existing alias-analysis path for inputs with unknown origins. The fast path avoids N copies of the operand list and N(N-1) pairwise checks for distinct allocation roots. ### Benchmark Measured with Release + Assertions `mlir-opt` on WSL2 (median total command runtime): | MemRefs | Baseline | Patched | Speedup | |---:|---:|---:|---:| | 512 | 0.518 s | 0.424 s | 1.22x | | 1024 | 0.785 s | 0.434 s | 1.81x | | 2048 | 1.851 s | 0.429 s | 4.31x | | 4096 | 6.414 s | 0.468 s | 13.69x | For an input with a `memref.copy` for each allocated buffer (nonempty transformed IR), 4096 memrefs improved from 6.128 s to 0.494 s (12.41x). ### Testing - Built baseline and patched `mlir-opt` successfully from LLVM revision `902c28258ab67d9095e5b8dbf51dce7d0c90d9a7`. - Passed the existing `buffer-deallocation-simplification.mlir` test. - Passed the new FileCheck cases for distinct allocation roots, aliased views sharing a root, unknown origins, and mixed cases. - Verified identical transformed IR between baseline and patched binaries for the tested cases. Fixes #219740. ### AI assistance ChatGPT assisted with the C++ implementation, regression tests, benchmark scripting, and preparation of this PR description. Baseline and modified `mlir-opt` binaries were built locally, and the reported FileCheck tests and benchmarks were executed on those binaries.