[AArch64] Recurse further into performUADDVCombine
authorDavid Green <david.green@arm.com>
Tue, 7 Feb 2023 19:14:06 +0000 (19:14 +0000)
committerDavid Green <david.green@arm.com>
Tue, 7 Feb 2023 19:14:06 +0000 (19:14 +0000)
commita2bd37230706ecb164c7ebe5cac66f0fd80b778d
treec746ab2e1c831be7e6fc7521c1b60a8fd5d6a982
parent3be1ae24fb17bf90cb3d8fc8e83ebd9d4523a594
[AArch64] Recurse further into performUADDVCombine

This adjusts the performUADDVCombine to recurse further into adds from
reductions, which can help produce more UADDLP when the order of lanes is
unimportant. It also prevents some regressions from subsequent patches.
llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
llvm/test/CodeGen/AArch64/vecreduce-add.ll