commit 41f6ad560f41cd43ee989897704457b595483f44
author David Benjamin <davidben@google.com> Sat Oct 28 01:26:50 2023
committer Crashpad LUCI CQ <crashpad-scoped@luci-project-accounts.iam.gserviceaccount.com> Mon Oct 30 19:31:22 2023
tree 5abca6c1dbdb99f19f07b6faeb7df44969b65821
parent 4a93d7f4c407fee2168ea23195d0e30fbfc1f90c

Fix crashpad tests under UBSan

These are slightly frustrating. First, when a struct is packed, some of
its fields may be underaligned. This is fine for direct access
(foo.bar), but if one takes the address if the field, this creates an
unaligned pointer. Dereferencing that pointer is then UB. (I'm not sure
if creating that pointer is UB.)

Crashpad seemingly doesn't do this, but it uses EXPECT_EQ from GTest.
EXPECT_EQ seems to internally take pointers to its arguments. I'm
guessing it binds them by const reference. This then trips UBSan. To
avoid this, we can copy the value into a temporary before passing to
EXPECT_EQ.

Second, the test to divide by 0 to trigger SIGFPE is undefined behavior.
The compiler is not actually obligated to trip SIGFPE. UBSan prints one
of its errors instead. Instead, since this file is only built on POSIX
anyway, use GCC inline assembly to do the division. That one is
well-defined.

Finally, casting a string to uint32_t* is undefined both by alignment
and by strict aliasing (although Chromium doesn't enable the latter).
Instead, type-punning should be done with memcpy.

Bug: chromium:1394755
Change-Id: I79108773a04ac26f5189e7b88a0acbf62eb4401d
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/4985905
Reviewed-by: Robert Sesek <rsesek@chromium.org>
Commit-Queue: David Benjamin <davidben@chromium.org>