[llvm] [llvm] Improve llvm.objectsize computation by computing GEP, alloca a… (PR #117849)

[via llvm-commits]

serge-sans-paille wrote:

Simpler reproducer:

```
unsigned char check_loopfilter_16y(unsigned i, int dir, long * res)
{
  unsigned char base1[32 + 16 * 16];
  int midoff = dir ? 4 * 16 : 4;
  int midoff_aligned = dir ? 4 * 16 : 16;
  unsigned char *buf1 = base1 + midoff_aligned;
  *res = __builtin_object_size((buf1 - midoff), 0);
  return base1[i];
}
```

Basically what happens here is that `buf1` is computed to be in `[base1 + 16, base1 + 4 * 16]` and `-midoff` in `[-4 * 16, -4]`. Because we have no data dependency, `__builtin_object_size` computes that its argument may be *before* `base1` (through `16 - 4*16`) and correctly detects it, but the corresponding action is probably not the best. please give me a day to make sure my local patch is correct, otherwise we'll revert. 

https://github.com/llvm/llvm-project/pull/117849