[llvm] [RFC] [msan] make MSan up to 20x faster on AMD CPUs (PR #171993)

[Thurston Dang via llvm-commits]

thurstond wrote:

MEM_TO_SHADOW in llvm/llvm-project/compiler-rt/lib/msan/msan.h would need to be updated as well.

Also, I think the kMemoryLayout might have some ambiguity with the end addresses of each region that didn't matter with the existing `0x500000000000` constant, but will matter with the modified constant e.g.,
* 0x010000000000 (app-1) ^ 0x500000000000 = 0x510000000000 (shadow-1)
* 0x010000000000 (app-1) ^ 0x500000000080 = 0x510000000080 (past the end of shadow-1)

https://github.com/llvm/llvm-project/pull/171993