llvm-reduce should remove lifetime.start/end when promoting alloca'd ptr to arg

[ManuelJBrito]

Lifetime.start/end in general should only be used for alloca'd pointers. When running the "operands-to-args" pass the lifetimes should be dropped.

define void @src() {
entry:
  %p = alloca i32, align 4
  call void @llvm.lifetime.start.p0(i64 4, ptr %p)
  store i32 10, ptr %p, align 4
  %a = load i32, ptr %p, align 4
  call void @llvm.lifetime.end.p0(i64 4, ptr %p)
  ret void
}

Running llvm-reduce --delta-passes=operands-to-args --test=test.sh src.ll

; ModuleID = 'src.ll'
source_filename = "src.ll"

define void @src(ptr %p) {
entry:
  %p1 = alloca i32, align 4
  call void @llvm.lifetime.start.p0(i64 4, ptr %p)
  store i32 10, ptr %p, align 4
  %a = load i32, ptr %p, align 4
  call void @llvm.lifetime.end.p0(i64 4, ptr %p)
  ret void
}

cc @nunoplopes @regehr

[arsenm]

I don't think I agree with this. It is not structurally required to remove the intrinsics. Reductions should make minimal structurally correct changes, the semantics are not so relevant (unless we introduce a preserve-semantics mode).

The LangRef does not state that it's invalid IR and the verifier does not reject it. Unless the IR verifier rules change, I think this should be left alone. If the use of the pointer does not matter, the operand use reductions on the intrinsic call should strip the argument use later

[nunoplopes]

The smaller the file, usually the better. Often the lifetime markers are irrelevant to bug repros.
IMO, having a pass in llvm-reduce that removes unused intrinsics is very useful.

[arsenm]

The smaller the file, usually the better. Often the lifetime markers are irrelevant to bug repros. IMO, having a pass in llvm-reduce that removes unused intrinsics is very useful.

Sure, they should eventually be removed as a consequence of the complete delta pass pipeline. I'm saying that operands-to-args need not be responsible for this, and these should be deleted later anyway