[PATCH] D156358: [AArch64] Do not unnecessarily spill LR because of @llvm.returnaddress

[Anatoly Trosinenko via Phabricator via llvm-commits]

atrosinenko added a comment.

> https://github.com/ahmedbougacha/llvm-project/commit/7924c7d75ae0015a9fd9786a580b10b2190bccc6
>
> @atrosinenko do you want to try out that patch, see if it helps?

Actually, this patch is inspired by the above commit (the commit mentioned in my first message is the same, it was just referenced through `apple/llvm-project`). Turned out, most of its features were previously reimplemented in mainline in a way similar to your implementation but surprisingly not preventing your patch from being applied. The optimization of `@llvm.returnaddress` lowering was among a few features of 7924c7d75a not yet existing in mainline (the second feature is Darwin-specific optimization). In 7924c7d75a rebased onto recent mainline commit, it seems to have a minor issue of emitting extra `mov`s sometimes, like this:

  _test_returnaddress_1:                  ; @test_returnaddress_1
  ; %bb.0:
          stp     x29, x30, [sp, #-16]!           ; 16-byte Folded Spill
          mov     x29, sp
          ldr     x8, [x29]
          ldr     x8, [x8, #8]
          mov     x0, x8
          xpaci   x0
          ldp     x29, x30, [sp], #16             ; 16-byte Folded Reload
          ret

Anyway, now I wonder whether we should not touch `@llvm.returnaddress` lowering at all and adjust register allocation somehow (RET_ReallyLR definition, getRegAllocationHints method, something else...).


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D156358/new/

https://reviews.llvm.org/D156358