[llvm] [SLP]Fix/improve getSpillCost analysis (PR #129258)

[Alexander Kornienko via llvm-commits]

alexfh wrote:

We've found a clang crash that seems to be introduced by this commit. The original build is using instrumented FDO + thinlto with a bunch of dependencies. I managed to get a standalone reproducer (hopefully, nothing important was lost during the automated reduction):
```
$ cat reduced.ll
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128-Fn32"
target triple = "aarch64-unknown-linux-gnu"

define void @_f(i64 %0) {
  switch i32 0, label %.lr.ph126 [
    i32 0, label %.loopexit35
    i32 1, label %.loopexit35
  ]

.loopexit46:                                      ; preds = %27
  %2 = load i64, ptr getelementptr inbounds nuw (i8, ptr null, i64 72), align 8
  %3 = or i64 %2, %21
  store i64 %3, ptr getelementptr inbounds nuw (i8, ptr null, i64 72), align 8
  %4 = load i64, ptr getelementptr inbounds nuw (i8, ptr null, i64 48), align 8
  %5 = or i64 %4, %22
  store i64 %5, ptr getelementptr inbounds nuw (i8, ptr null, i64 48), align 8
  %6 = load i64, ptr getelementptr inbounds nuw (i8, ptr null, i64 64), align 8
  %7 = or i64 %6, %23
  store i64 %7, ptr getelementptr inbounds nuw (i8, ptr null, i64 64), align 8
  %8 = load i64, ptr getelementptr inbounds nuw (i8, ptr null, i64 56), align 8
  %9 = or i64 %8, %24
  store i64 %9, ptr getelementptr inbounds nuw (i8, ptr null, i64 56), align 8
  br label %.loopexit35

.lr.ph126:                                        ; preds = %27, %1
  %10 = phi i64 [ %0, %27 ], [ 0, %1 ]
  %11 = phi i64 [ 0, %27 ], [ 0, %1 ]
  %12 = phi i64 [ 0, %27 ], [ 0, %1 ]
  %13 = phi i64 [ 0, %27 ], [ 0, %1 ]
  br i1 false, label %.loopexit39, label %.lr.ph

.lr.ph:                                           ; preds = %16, %.lr.ph126
  %14 = phi i64 [ %17, %16 ], [ 0, %.lr.ph126 ]
  %15 = phi i64 [ 0, %16 ], [ 0, %.lr.ph126 ]
  br i1 false, label %.loopexit39, label %16

16:                                               ; preds = %.lr.ph
  %17 = add i64 0, 0
  br i1 false, label %.lr.ph, label %.loopexit39

.loopexit39:                                      ; preds = %16, %.lr.ph, %.lr.ph126
  %18 = phi i64 [ 0, %.lr.ph126 ], [ %15, %.lr.ph ], [ %14, %16 ]
  %19 = phi i64 [ 0, %.lr.ph126 ], [ %14, %.lr.ph ], [ %17, %16 ]
  %20 = phi i64 [ 0, %.lr.ph126 ], [ %14, %.lr.ph ], [ %14, %16 ]
  %21 = or i64 %19, %13
  %22 = or i64 %20, %12
  %23 = or i64 %18, %11
  %24 = or i64 %18, %10
  br label %27

25:                                               ; No predecessors!
  %26 = call ptr null(i64 0)
  br label %27

27:                                               ; preds = %25, %.loopexit39
  br i1 false, label %.lr.ph126, label %.loopexit46

.loopexit35:                                      ; preds = %.loopexit46, %1, %1
  ret void
}

; uselistorder directives
uselistorder ptr null, { 0, 4, 3, 2, 1 }
$ opt -passes slp-vectorizer reduced.ll -o /dev/null
Instruction does not dominate all uses!
  %11 = insertelement <2 x i64> %10, i64 0, i32 1
  %8 = shufflevector <2 x i64> %11, <2 x i64> poison, <2 x i32> <i32 1, i32 poison>
LLVM ERROR: Broken module found, compilation aborted!
```

https://github.com/llvm/llvm-project/pull/129258