flagos-ai/flir：FlagTree IR is forked from microsoft/triton-shared, which is a Shared Middle-Layer for Triton Compilation. It is used for FlagTree.

tsingmicro

[Tsingmicro] sync tsingmicro’s flir patches for LLVM22 support (#69)

sync tsingmicro’s flir patches for LLVM22 support

[Fix] fix MaskAnalysis dropping sge-against-0 mask with non-zero lower bound


parseCmp assumed cmpi sge x, 0 always evaluates to true, which drops the
mask when x has a statically negative or dynamic lower bound (e.g. the
(offs - n) >= 0 pattern in causal_conv1d). The load/store was then
rewritten as a full-range DMA that reads below the base pointer and
triggers the hardware DMA OOB checker. Bail out instead so the unstructured
gather/scatter path preserves the mask.
Add masked_sge_shifted.mlir regression test; update addptr_cmpge.mlir for
the new fallback behavior of the dynamic-lower-bound load case.
Change-Id: I3d866c10823f9f1ddeed2885eb77932779560ddd

guard LLVM22-only builder APIs to keep <LLVM22 builds compiling


Co-authored-by: yinle yinle@dev.com