| Files | Last commit | Last update |
|---|---|---|
[AscendNPU-IR] add llvm-project patch: bufferization.to_buffer for bytecode reading LLVM 22 emits bufferization.to_buffer as a registered op with native properties (kHasProperties; inherent attr UnitAttr:$read_only). LLVM 19's BytecodeReader does not know this op, so PropertiesSectionReader::read (BytecodeReader.cpp:1122) falls into the 'unregistered -> readAttribute' branch and tries to decode native-properties bytes as an attribute index, yielding 'invalid Attribute index: 9' and crashing. Add a Bufferization_ToBufferOp definition adapted to LLVM 19: operand AnyTensor:$tensor, result AnyRankedOrUnrankedMemRef:$buffer, inherent attr UnitAttr:$read_only, and the upstream assembly format. Declaring the inherent attr makes hasProperties() true, so mlir-tblgen auto-attaches BytecodeOpInterface::Trait (OpDefinitionsGen.cpp:3843-3844) and generates readProperties/writeProperties for read_only -- matching LLVM 22's properties encoding, so the reader takes the iface->readProperties path instead of crashing. (LLVM 22's Bufferization_TensorLikeTypeInterface / Bufferization_BufferLikeTypeInterface / Bufferization_TensorAndBufferMatch do not exist in LLVM 19, so AnyTensor/AnyRankedOrUnrankedMemRef are used and the match-trait is dropped.) Patch applied to third-party/llvm-project at build time via build-tools/apply_patches.sh. | 3 days ago | |
[AscendNPU-IR] fix: fix build check | 2 months ago | |
delete apply-patches | 8 days ago | |
Update build.sh to optimize build step | 2 months ago |
| Files | Last commit | Last update |
|---|---|---|
| 3 days ago | ||
| 2 months ago | ||
| 8 days ago | ||
| 2 months ago |