[NVPTX] MIR optimizations incorrectly deleting return value register

[jhuber6]

We are currently failing to build any of the NVPTX libc test suite. The following IR fails to validate with ptxas when run through the backend with optimizations https://godbolt.org/z/ahro9n3KE.

target triple = "nvptx64-nvidia-cuda"

declare ptr @bar(i64)

declare i64 @baz()

define ptr @foo(i1 %cond) {
entry:
  %call = call i64 @baz()
  br i1 %cond, label %common.ret, label %bb

common.ret:
  ret ptr null

bb:
  %tmp = call ptr @bar(i64 %call)
  br label %common.ret
}

$ llc reduced.ll -mcpu=sm_89 -o reduced.s
$ ptxas -m64 -v --gpu-name sm_89 reduced.s -c
ptxas reduced.s, line 39; error   : Illegal operand type to instruction 'ld'
ptxas reduced.s, line 39; error   : Unknown symbol 'retval0'

If you look at the output PTX it just straight up deletes the reval0 register without cleaning up its users. If you compile without optimizations the node is present, so it occurs due to some MIR optimization.

[Artem-B]

It must be indeed a recent change. The issue is not reproducible with 30705c5

[AlexMaclean]

Apologies, it seems like machine sink is sinking a load of the retval past the end of the call sequence. I'm working on a fix now.

[AlexMaclean]

@jhuber6 Just wanted to confirm that the NVPTX libc test suite is back to passing after my recent change?

[jhuber6]

@jhuber6 Just wanted to confirm that the NVPTX libc test suite is back to passing after my recent change?

Yep, worked last time I tested it. Benefits of having a bunch of random unit tests I suppose. https://lab.llvm.org/buildbot/#/builders/69