[llvm] [VPlan] First step towards VPlan cost modeling. (PR #92555)

[Harald van Dijk via llvm-commits]

hvdijk wrote:

I'm also hitting the same assert, reduced test courtesy of bugpoint is:

```llvm
target datalayout = "e-m:e-p:64:64-i64:64-i128:128-n32:64-S128"
target triple = "riscv64-unknown-unknown-elf"

define void @foo() {
loopIR.i:
  br label %loopIR2.i

loopIR2.i:
  %0 = phi i64 [ 0, %loopIR.i ], [ %1, %loopIR2.i ]
  %.pre = load i32, ptr addrspace(1) null, align 4
  store i32 %.pre, ptr addrspace(1) poison, align 4
  %1 = add nuw i64 %0, 1
  %exitcond1.not = icmp eq i64 %1, poison
  br i1 %exitcond1.not, label %exitIR4.i, label %loopIR2.i

exitIR4.i:
  ret void
}
```

```
$ opt -mattr=+v -passes=loop-vectorize -S -o - bug.ll
```

I'm not sure whether the initial testing was with or without the already-pushed fixes, I will re-test with current `main`.

https://github.com/llvm/llvm-project/pull/92555