[clang] [llvm] [AMDGPU] Use a general form of intrinsic for tensor load/store (PR #182334)

Krzysztof Drewniak via cfe-commits cfe-commits at lists.llvm.org
Fri Feb 20 10:57:41 PST 2026


krzysz00 wrote:

@changpeng https://github.com/llvm/llvm-project/blob/be7267002357a3955d225da84d7fcf6c7e3540b3/mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td#L1275-L1331 is the relevant code - you'll want to delete the "d2" class and update the load class to give it the extra parameter

>From there, https://github.com/llvm/llvm-project/blob/ef55e5938936fa51b88b2a8008c8a2821ec51dd5/mlir/test/Dialect/LLVMIR/rocdl.mlir#L814 are the parse/print tests, https://github.com/llvm/llvm-project/blob/ef55e5938936fa51b88b2a8008c8a2821ec51dd5/mlir/test/Target/LLVMIR/rocdl.mlir#L1272 are the translation tests - the updates should be fairly mechanical

To match the extra parameter, go ahead and update https://github.com/llvm/llvm-project/blob/4c380cfe3fbf20f470de32ed52756417bdd1eaaf/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp#L3697 to create constant 0s and pass them in as a 5th argument and add a comment that that'll move into the TDM descriptor type once it becomes relevant for future use.

https://github.com/llvm/llvm-project/blob/47d420a2c8e0f9edbdf58ce8b8dcc7d6743b4191/mlir/test/Conversion/AMDGPUToROCDL/gfx1250.mlir#L728 will need updating to account for that extra constant of splat 0s.

https://github.com/llvm/llvm-project/pull/182334


More information about the cfe-commits mailing list