[StackProtector] don't check stack protector before calling nounwind functions
authorNick Desaulniers <ndesaulniers@google.com>
Thu, 13 Apr 2023 16:30:19 +0000 (09:30 -0700)
committerNick Desaulniers <ndesaulniers@google.com>
Thu, 13 Apr 2023 16:37:06 +0000 (09:37 -0700)
commitfc4494dffa5422b2be5442c235554e76bed79c8a
treeee82fbbf5701e5bba68184fa625c2e6ec410a262
parentf0e65291339f1b27d11a1ce2e908a614b7f4cb8c
[StackProtector] don't check stack protector before calling nounwind functions

https://reviews.llvm.org/rGd656ae28095726830f9beb8dbd4d69f5144ef821
introduced a additional checks before calling noreturn functions in
response to this security paper related to Catch Handler Oriented
Programming (CHOP):
https://download.vusec.net/papers/chop_ndss23.pdf
See also:
https://bugs.chromium.org/p/llvm/issues/detail?id=30

This causes stack canaries to be inserted in C code which was
unexpected; we noticed certain Linux kernel trees stopped booting after
this (in functions trying to initialize the stack canary itself).
https://github.com/ClangBuiltLinux/linux/issues/1815

There is no point checking the stack canary like this when exceptions
are disabled (-fno-exceptions or function is marked noexcept) or for C
code.  The GCC patch for this issue does something similar:
https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=a25982ada523689c8745d7fb4b1b93c8f5dab2e7

Android measured a 2% regression in RSS as a result of d656ae280957 and
undid it globally:
https://android-review.googlesource.com/c/platform/build/soong/+/2524336

Reviewed By: xiangzhangllvm

Differential Revision: https://reviews.llvm.org/D147975
llvm/lib/CodeGen/StackProtector.cpp
llvm/test/CodeGen/X86/stack-protector-2.ll
llvm/test/CodeGen/X86/stack-protector-recursively.ll