[llvm] [AMDGPU] Move kernarg preload logic to AMDGPU Attributor (PR #123547)

[Austin Kerbow via llvm-commits]

https://github.com/kerbowa commented:

>> In one case this leads to one extra argument being preloaded in a test. ISel correctly identifies this opportunity even when the IR pass previously missed it.

> Which case is this? Getting the exact number of arguments at all points should not be difficult

`half_v7bfloat_kernel_preload_arg` as an example. The difference between the IR tracking of this and ISel is that ISel would always allocate in-order whereas in IR we would sometimes see if we could fit a randomly located argument which had bugs around arguments in previously allocated SGPRs and also the alignment of the hidden arguments. Looking just at the offsets in the same way the HSA metadata is emitted avoids all of this since we don't need to track the specific registers that arguments are preloaded to in IR to determine what ISel will be able to preload.

https://github.com/llvm/llvm-project/pull/123547