https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104971
Bug ID: 104971
Summary: Optimisation for __builtin_ia32_readeflags corrupts
the stack
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: andrew.cooper3 at citrix dot com
Target Milestone: ---
Full example: https://godbolt.org/z/xGq3c4Mnc
Given:
int broken(void)
{
int fl = __builtin_ia32_readeflags_u64();
}
gcc -O2 generates:
broken:
pushfq
ret
Which is going explode very quickly.
Code generation appears to be safe without optimisation, but even -O alone is
enough to create problems.
At a guess, the optimiser has concluded that the result is unused, drops the
`pop %reg`, but fails to also drop the `pushf` too.
Looking through history on Godbolt, it appears that GCC 4.9 (which introduced
this builtin) has correct optimised code generation, and it regressed between
4.9 and 5.1.