Maybe mis-optimization: Move const array initialized inplace to unname_addr const

[tesuji]

This happens in real word code in rust: rust-lang/rust#73825
Clang and zig generates better IR so llvm could optimize const non-mutated arrays to .rodata.

Godbolt link of LLVM-IR: https://godbolt.org/z/PGdYYx54r
Consider this:

define i32 @square(i64 noundef %x) local_unnamed_addr #0 {
  %base = alloca [256 x i8], align 16
  call void @llvm.lifetime.start.p0(i64 256, ptr nonnull %base)
  store <4 x i32> <i32 67, i32 754, i32 860, i32 559>, ptr %base, align 16
  %1 = getelementptr inbounds [64 x i32], ptr %base, i64 0, i64 4
  store <4 x i32> <i32 368, i32 870, i32 548, i32 972>, ptr %1, align 16
...
  %_3 = and i64 %x, 63
  %16 = getelementptr inbounds [64 x i32], ptr %base, i64 0, i64 %_3
  %_0 = load i32, ptr %16, align 4
  call void @llvm.lifetime.end.p0(i64 256, ptr nonnull %base)
  ret i32 %_0
}

I expect the above IR optimized to:

@square.base = private unnamed_addr constant <{ [256 x i8] }> <{ [256 x i8] c"C\00\00\00\F2\02\00\00\\\03\00\00/\02\00\00p\01\00\00f\03\00\00$\02\00\00\CC\03\00\00\8D\00\00\00\DB\02\00\00_\01\00\00\98\02\00\00 \00\00\00\04\00\00\00\E4\03\00\00\E5\02\00\00\CB\00\00\00$\01\00\00\ED\00\00\00\E0\01\00\00\97\00\00\00\AC\03\00\00\09\03\00\00\1C\02\00\00\8F\00\00\00K\02\00\00\EB\02\00\00A\00\00\00\98\00\00\00\05\02\00\00r\03\00\00p\03\00\00\C8\02\00\00S\02\00\00r\01\00\00\85\03\00\00\ED\00\00\005\00\00\00\15\03\00\00\11\03\00\00\90\03\00\00\8A\02\00\00\80\03\00\00o\01\00\00<\01\00\00\88\01\00\00>\00\00\00\D9\01\00\00\A3\02\00\00\B3\02\00\00\19\01\00\00\C0\00\00\00\BD\01\00\00\CA\03\00\00\E1\00\00\00\A9\01\00\00t\02\00\00D\01\00\00B\01\00\00\CE\00\00\00\90\03\00\00c\03\00\00\CE\01\00\00\\\00\00\00" }>, align 4

define i32 @square(i64 noundef %x) unnamed_addr #0 {
  %_3 = and i64 %x, 63
  %0 = getelementptr inbounds [64 x i32], ptr @square.base, i64 0, i64 %_3,
  %_0 = load i32, ptr %0, align 4, !noundef !11
  ret i32 %_0
}

I know the IR that rustc generates is poor. I try to improve it but some people wonder if LLVM could do its marvelous job again with current IR.