Segmentation fault in SPEChpc 2021 sph_exa_t with LTO enabled

I built sph_exa_t using clang++ 21.1.8.
The options applied were:
-O3 -flto -mcpu=native

When I built and ran with the above options, a SEGV occurred.
The SEGV was found to occur within the _Tr::allocate function, which is part of the std::vector implementation.

- /usr/include/c++/11/bits/stl_vector.h
``` C++
pointer
_M_allocate(size_t __n)
{
typedef __gnu_cxx::__alloc_traits<_Tp_alloc_type> _Tr;
return __n != 0 ? _Tr::allocate(_M_impl, __n) : pointer();
}
```

My analysis suggests that this is likely caused by the following IR changes introduced by SimplifyCFG during LTO (Link-Time Optimization). I confirmed this by using -Wl,-mllvm,-print-before=simplifycfg and -Wl,-mllvm,-print-after=simplifycfg.

- IR code before SimplifyCFG application:
(Check with -Wl,-mllvm,-print-before=simplifycfg)
```
8: ; preds = %0
%9 = icmp eq i32 %4, 0
br i1 %9, label %18, label %10

10: ; preds = %8
%11 = shl nuw nsw i64 %5, 2
%12 = call noalias noundef nonnull ptr @_Znwm(i64 noundef %11) #7
%13 = getelementptr inbounds nuw i32, ptr %12, i64 %5
store i32 0, ptr %12, align 4, !tbaa !8
%14 = icmp eq i32 %4, 1
br i1 %14, label %18, label %15
```
- IR code after SimplifyCFG application:
(Check with -Wl,-mllvm,-print-after=simplifycfg)
```
7: ; preds = %0
%8 = zext nneg i32 %4 to i64
%9 = icmp ne i32 %4, 0
call void @llvm.assume(i1 %9)
%10 = shl nuw nsw i64 %8, 2
%11 = call noalias noundef nonnull ptr @_Znwm(i64 noundef %10) #8
store i32 0, ptr %11, align 4, !tbaa !8
%12 = icmp eq i32 %4, 1
br i1 %12, label %16, label %13
```

I have attached a reproducer code snippet that is a significantly reduced and modified version of the actual code.
[src_and_script.zip](https://github.com/user-attachments/files/24706013/src_and_script.zip)

Start with the attached src_and_script.zip and reproduce the crash using clang++ 21.1.8 with -O3 -flto -mcpu=native. Compare the SimplifyCFG IR before and after the pass using the two -Wl,-mllvm print options; done means the reduced reproducer no longer segfaults and the incorrect transformation is addressed.

Written by the indexing model from the issue text.