[PATCH 10/12] aarch64: Add pac-ret support to asm files

[Adhemerval Zanella]

On 11/05/2020 05:27, Szabolcs Nagy wrote:
> The 05/08/2020 13:59, Adhemerval Zanella wrote:
>> On 30/04/2020 14:44, Szabolcs Nagy wrote:
>>> From de8968ed58686c26391de8343184a1283bb5e305 Mon Sep 17 00:00:00 2001
>>> From: Szabolcs Nagy <szabolcs.nagy@arm.com>
>>> Date: Wed, 29 Apr 2020 11:49:20 +0100
>>> Subject: [PATCH 10/12] aarch64: Add pac-ret support to asm files
>>>
>>> This patch unconditionally enables pac-ret in asm files.
>>>
>>> TODO: This will need configure checks, cannot be done
>>> unconditionally because we cannot guarantee pac-ret
>>> compatibility (e.g. libgcc unwinder had no support for
>>> it before gcc-7 and newer libgcc had bugs that could
>>> cause unwind crash when pac-ret and non-pac-ret stack
>>> frames are mixed)
>>
>> Which gcc version does it work correctly? Would it be a check against
>> a specific gcc version or could it be a configure to see if libgcc
>> provides the expected fixes? 
>>
>> The gcc-7 branch is now closed, so maybe we could assume gcc-8 and 
>> have the required fixes backported?
> 
> the version number is not useful because the fix
> can be backported (even to closed branches, just
> not upstream).
> 
> but gcc 10.1 has no known bug.
> (except if -pg and __builtin_return_address
> handling changes, i consider current -pg
> broken with pac-ret, but we might decide to
> fix it up in _mcount instead of the compiler)

Ack.

> 
>>
>> In any case I think we should add a configure check based on compiler
>> -mbranch-protection= options used to enable/disable ENABLE_PAC_RET.
> 
> yeah i changed the configury bits in v2, i didnt
> do it in v1 because it's quite painful: gcc
> does not set a macro to test for it
> 

Yes, such checks requires to build a code snippet and parse for specific
instructions patterns.