[clang] Bfi precision (PR #66285)

[Matthias Braun via cfe-commits]

MatzeB wrote:

Some very ad-hoc benchmarking. Of clang compilation speed (measured in instructions as reported by valgrind/callgrind which I think somewhat matches the setup of nikic) compiling `sqlite3` of CTMark:

Old-BFI (this PR reverted), New-BFI (this PR applied), no-cold (cold-callsite-rel-freq set to 0 to disable `InlineCostCallAnalyzer::isColdCallSite` behavior for non-PGO builds:

```
Old-BFI:          size:  59,802,772  (baseline)  insn: 2,812,833,144  (baseline)
New-BFI:          size:  60,247,076  +0.74%      insn: 2,818,639,641  +0.21%
Old-BFI, no-cold: size:  60,773,988  +1.62%      insn: 2,806,521,932  -0.22%
New-BFI, no-cold: size:  60,741,700  +1.57%      insn: 2,803,489,298  -0.33%
```

I could benchmark more with llvm-test-suite and put up a PR to disable cold-callsite-rel-freq by default if people support this direction...

https://github.com/llvm/llvm-project/pull/66285