This is to teach tailcallelim transformation not to mark
llvm.stackrestore with tail-call, as the intrinsic call can modify
unescaped allocas from the caller.
This fixes a problem found with our downstream testing. The problem can
also be shown when running the test case
llvm/test/Transforms/MemCpyOpt/stackrestore with passes=”tailcallelim,
memcpyopt”.