[Bug target/86832] [8/9 Regression] GCC v8.2.0 tries to use native TLS with -fstack-protector-strong on Windows (mingw-w64)

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86832

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |uros at gcc dot gnu.org

--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I believe this is caused by the PR81708 changes.
While i386 defaulted to SSP_TLS rather than SSP_GLOBAL on everything but
Android,
the -mstack-protector-guard= switch controlled pretty much whether the i386.md
special stack protector patterns are used (if tls) or whether generic code is
used (global).  These special stack protector patterns did one thing if
TARGET_THREAD_SSP_OFFSET macro was defined (only defined on glibc targets) -
code like:
        movq    %fs:40, %rax
        movq    %rax, -8(%rbp)
        xorl    %eax, %eax
in the prologue and
        movq    -8(%rbp), %rdx
        xorq    %fs:40, %rdx
        je      .L4
in the epilogue.  If TARGET_THREAD_SSP_OFFSET macro wasn't defined, it would do
instead:
        movq    .refptr.__stack_chk_guard(%rip), %rax
        movq    (%rax), %rcx
        movq    %rcx, -8(%rbp)
        xorl    %ecx, %ecx
and
        movq    .refptr.__stack_chk_guard(%rip), %rdx
        movq    -8(%rbp), %rcx
        xorq    (%rdx), %rcx
        je      .L4
(this is taken from 7.x cross to mingw).
Finally, for Android or when -mstack-protector-guard=global was used, it
emitted:
        movq    __stack_chk_guard(%rip), %rax
        movq    %rax, -8(%rbp)
and
        movq    __stack_chk_guard(%rip), %rdx
        cmpq    %rdx, %rcx
        je      .L4
Note, apart from OS specific details, those =global sequences are similar to
the =tls ones when TARGET_THREAD_SSP_OFFSET is not defined, the main difference
is that the =tls ones are more secure as they clear registers containing the
guard as quickly as possible.  The PR81708 changes dropped the non-tls special
stack_protector_* patterns from i386.md and now =tls implies really tls, but
the default remained, so mingw32 or darwin still default to tls and just use 0
offset by default.

So, this patch changes the default for mingw32, darwin and everything else
except gnu-user*.h to be =global, and just forces those special i386.md more
secure patterns unconditionally (slightly changing the generated code on
Android, but it is one extra insn in prologue and one fewer in the epilogue).