[BUG] benchmark::DoNotOptimize does not prevent a reference from being optimized away.

**Describe the bug**
`benchmark::DoNotOptimize()` does not prevent a reference from being optimized away.

**System**
OS: Ubuntu 20.04 LTS, Compiler explorer
Compiler: GCC 9.04, GCC trunk, Clang trunk

**To reproduce**
Compiler Explorer that works: https://godbolt.org/z/c3d5sc8s3
Compiler Explorer that does not: https://godbolt.org/z/3bMMvaecr

Alternatively, compile the following code
```c++
#include
#include
static void bm_traverse(benchmark::State& state) {
const size_t N = 500;
std::vector vec;
for (size_t j = 0; j < N; ++j) vec.push_back(j);
for (auto _ : state) {
for (auto& value : vec) {
asm("nop"); // just to locate asm point faster
benchmark::DoNotOptimize(value);
}
benchmark::DoNotOptimize(vec);
}
state.SetComplexityN(N);
}
BENCHMARK(bm_traverse);
BENCHMARK_MAIN();
```

**Expected behavior**
DoNotOptimize should have preserved the loop read. I understand that the reference is basically a pointer but users typically would expect that the reference represents the actual lvalue.

**Screenshots**

With the "&" reference in the inner loop, a vec load is not generated.
```asm
.LBB0_31: # Parent Loop BB0_6 Depth=1
nop
addq $4, %rax
cmpq %rcx, %rax
jne .LBB0_31
```
If you remove the reference the vec load is generated.
```asm
.LBB0_31: # Parent Loop BB0_6 Depth=1
movl (%rax), %edx # <<<< HERE
movl %edx, 12(%rsp)
nop
addq $4, %rax
cmpq %rcx, %rax
jne .LBB0_31
```

**Additional context**

I understand this is a matter of semantics that might have been discussed in the past. In this case, I'd just like confirmation of this behavior.