projects / platform / upstream / llvm.git / commit
? search:
summary | shortlog | log | commit | commitdiff | tree
(parent: 9b704d3) | patch
Revert "[MLIR] Update Vector To LLVM conversion to be aware of assume_alignment"
authorStephen Neuendorffer <stephen.neuendorffer@xilinx.com>
Tue, 30 Nov 2021 22:37:30 +0000 (14:37 -0800)
committerStephen Neuendorffer <stephen.neuendorffer@xilinx.com>
Tue, 30 Nov 2021 23:18:22 +0000 (15:18 -0800)
commit73863648892ee7063c7fd4e658d7614fd721504a
tree39730a3ee2eed6394100abbf6d0adb45b9ca7b06tree | snapshot
parent9b704d31b54a61615e7d172a1ed040bff7423e9bcommit | diff
Revert "[MLIR] Update Vector To LLVM conversion to be aware of assume_alignment"

This reverts commit 29a50c5864ddab283c1ff38694fb5926ce37b39a.

After LLVM lowering, the original patch incorrectly moved alignment
information across an unconstrained GEP operation.  This is only correct
for some index offsets in the GEP.  It seems that the best approach is,
in fact, to rely on LLVM to propagate information from the llvm.assume()
to users.

Thanks to Thomas Raoux for catching this.
mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp diff | blob | history
mlir/test/Conversion/VectorToLLVM/vector-to-llvm.mlir diff | blob | history
Domain: System / Toolchain;