[llvm] [AArch64] Remove superfluous sxtw in peephole opt (PR #96293)

[Krasimir Georgiev via llvm-commits]

krasimirgg wrote:

I rolled back the patch yesterday, sorry for not leaving a comment about that here in a timely manner.

We saw some issues with this in some halide-generated code.

It appears that after this, llc segfaults on the following code:
```
target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128-Fn32"
target triple = "aarch64--linux-eabi"

define i32 @f(i32 %0) {
entry:
  %1 = sext i32 %0 to i64
  br label %A

A:
  %2 = trunc i64 %1 to i32
  %a69.us = sub i32 0, %2
  %a69.us.fr = freeze i32 %a69.us
  %3 = zext i32 %a69.us.fr to i64
  br label %B

B:
  %t = icmp eq i64 0, %3
  br i1 %t, label %A, label %B
}
```

`assert.h assertion failed at .../llvm/lib/CodeGen/LiveVariables.cpp:159 in void llvm::LiveVariables::HandleVirtRegUse(Register, MachineBasicBlock *, MachineInstr &): MRI->getVRegDef(Reg) && "Register use before def!"`

https://github.com/llvm/llvm-project/pull/96293