https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92038
Bug ID: 92038
Summary: Extremely inefficient x86_64 code for trivally
copyable types passed in registers.
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: maxim.yegorushkin at gmail dot com
Target Milestone: ---
The following code:
#include <variant>
void f(std::variant<int, int>);
void g() { f({}); }
When compiled with `g++-9.2 -std=gnu++17 -O3 -march=skylake` generates the
following assembly:
g():
mov DWORD PTR [rsp-16], 0
mov BYTE PTR [rsp-12], 0
mov rdi, QWORD PTR [rsp-16]
jmp f(std::variant<int, int>)
Which is rather poor: unnecessary memory stores; dependency of rdi on the value
of 3 bytes of padding at [rsp-11], [rsp-10], [rsp-9] which are unset, which may
prevent store-to-load forwarding.
`clang++-8.0 -std=gnu++17 -O3 -march=skylake` generates the expected assembly:
g():
xor edi, edi
jmp f(std::variant<int, int>)
