[compiler-rt] [TSan] Clarify and enforce shadow end alignment (PR #144648)

[Marco Elver via llvm-commits]

melver wrote:

> > Consider reverting this commit, and then re-committing with the fix.
> 
> After debugging, I found that it is a behavior difference of `ReleaseMemoryPagesToOS` between IOS and Linux.
> 
> In IOS, it executes `madvise(beg, end, MADV_FREE)`, which tags the relevant pages as 'FREE' and does not release them immediately. In Linux, it executes `madvise(beg, end, MADV_DONTNEED)`, releasing the pages immediately and setting them to 0.
> 
> Therefore, in IOS, ` __tsan_read1(&buf[0]);` still raises a race error even after `munmap` 'clears' the relevant shadow pages.
> 
> I think we need to change the test case to one of the following fixes:
> 
> 1. Directly verify the page status rather than expect `__tsan_read1` not to raise an error (For an end-to-end testcase, this is too detailed).
> 2. Turn off the test case `munmap_clear_shadow.c`, because TSan does not provide a direct API query whether the shadow pages are cleared.
> 
> Do you think which fix is better, or can you give a better suggestion? @melver
> 
> After confirming the fix, I will propose another Reland PR. Or for such a simple testcase-only fix, should I just fix-forward without revert?

Turning off the test case is unacceptable - that'd be a regression.

Could you explain why this commit introduced the regression? 

https://github.com/llvm/llvm-project/pull/144648