AscendNPU-IR/bishengir/lib/Transforms · Ascend/AscendNPU-IR

Qqq3646[Huawei][AscendNPU IR] Debugging Capabilities: UB Probe: human readable

文件	最后提交记录	最后更新时间
AffineMinMaxValueBounds.cpp	[AscendNPU-IR] fix: replace llvm_unreachable with report_fatal_error	27 天前
CMakeLists.txt	[Tools] Support pass-scoped full-module IR injection Enable pass ID based injection in retriable pipelines so cross-version IR can replace the complete module at an exact compilation boundary. Co-authored-by: Cursor <cursoragent@cursor.com>	26 天前
CanonicalizeModule.cpp	[AscendNPU IR] Add HFusion Co-authored-by: chenxiangting<chenxiangting@huawei.com>	10 个月前
DeadFunctionElimination.cpp	[AscendNPU IR] Add HFusion Co-authored-by: chenxiangting<chenxiangting@huawei.com>	10 个月前
DebugMemory.cpp	[Huawei][AscendNPU IR] Debugging Capabilities: UB Probe: human readable	21 天前
EliminateSingleIterationScfFor.cpp	[Huawei][AscendNPU IR] migrate EliminateSingleIterationScfFor from A5 Migrate the global EliminateSingleIterationScfFor pass and its AffineMinMaxValueBounds helper from A5. The pass eliminates scf.for loops proven to execute exactly once using trip-count, ValueBounds, and affine min/max interval reasoning. A5 source commits: 48dc4bf11 ([Huawei][AscendNPUIR] deduce single loop and eliminate); 17c135eb6, 7ea38fdc4, d14b5d361 for subsequent func::FuncOp/affine-bound adjustments. A5 test commits: 48dc4bf11 for eliminate-single-iteration-scf-for-cases.mlir; 48dc4bf11, 17c135eb6, 7ea38fdc4, d14b5d361 for affine-min-max-value-bounds-coverage.mlir. Verification: ninja; ./bin/llvm-lit ../bishengir/test (baseline failure only: bishengir-compile/save-temps.mlir).	1 个月前
ExtendedCanonicalizer.cpp	[Huawei][AscendNPU IR] Work around A5-specific transpose folding logic The upstream FoldTransposeWithTranspose pattern was modified to avoid folding transposes that should later be decomposed for A5. However, this behavior is not suitable for A2/A3, where transpose folding should remain unconditional. Add a local copy of the original FoldTransposeWithTranspose pattern and use it for non-Ascend950 targets to preserve the previous behavior. Signed-off-by: kiansitov <iansitov.kirill@huawei.com>	1 个月前
FuseReductionIntoLoop.cpp	[Huawei][AscendNPU IR] migrate FuseReductionIntoLoop from A5 Migrate the global FuseReductionIntoLoop pass from A5. The pass fuses post-loop linalg.reduce users into scf.for loop bodies to carry lower-rank accumulators directly. A5 source commits: cdea94c83 ([Huawei][AscendNPU IR] Transform 2D to 1D reduction by fusing into loop). A5 test commits: cdea94c83 for bishengir/test/Transforms/fuse-reduction-into-loop.mlir. Verification: ninja -C build; build/bin/llvm-lit -sv bishengir/test/Transforms/fuse-reduction-into-loop.mlir.	1 个月前
InjectIR.cpp	[Huawei][AscendNPU IR] migrate InjectIR: apply function type mismatch fix from A5 Apply A5 fix for IR injection: update function type when mismatch detected. Test: build passes, llvm-lit shows no new failures. Co-Authored-By: Claude <noreply@anthropic.com>	1 个月前
InjectIRInstrumentation.cpp	[Tools] Normalize IR injection source line endings Remove CRLF artifacts retained by the legacy migration branch so whitespace checks pass on the replacement PR. Co-authored-by: Cursor <cursoragent@cursor.com>	26 天前
LowerToCPUBackend.cpp	[AscendNPU-IR] fix: replace llvm_unreachable with report_fatal_error	27 天前
Transforms.cpp	[AscendNPU-IR] fix regbase llvm - fix: Add A5 conditional compilation for ConstantIntOp in regbase/HIVMToStandard.cpp	1 个月前