gcc versions tested (all have this problem on both architectures):
x86_64
3.4.6 4.1.1 4.1.2
i686
3.4.6 4.1.1
Reduced testcase: main.cpp
---------------
struct o {
virtual void a() { }
};
int main(int,char**) {
new o();
return 0;
}
----------------
When a() is removed or set as non-virtual, problem disappears
Commandline to fail compilation:
i686:
g++ main.cpp -o main -O -fomit-frame-pointer -march=athlon -masm=intel
x86_64:
g++ main.cpp -o main -O -masm=intel
-march of athlon, athlon-4, k8 fails
without -masm the code is compiled, but its behavior is undefined (although
working)
---------------
In assembler output, the following line appears with AT&T syntax:
rep ; ret
and is accepted. (two independendend instructions, although meaningless)
With Intel syntax, generated code:
rep ret
is not accepted. (meaningless prefix)
---------------
According to Intel IA-32 documentation:
"The behavior of the REP prefix is undefined when used with non-string
instructions."
AMD AMD64 documentation:
"The (rep/repe) prefixes should only be used with such (lods,stos,movs,...)
string instructions."
---------------
I could post exact error messages and outputs of gcc -v on all tested systems
if needed (I have already done so, but the output is very long to past here)
------------
Conlusion:
Generates code non-conforming to Intel and AMD documentation. Unlikely, code
would not work on future processors. With AT&T syntax, asm code is accepted.
With Intel syntax, it is rejected.
The pair 'rep ; retn' should never appear as it is undefined when executed by
CPU. (and it most likely gets generated nonintentionally)
--
Summary: generating non-conforming "rep retn" with -match=athlon,
k8 ; with -masm=intel fails
Product: gcc
Version: 4.1.2
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: zsojka at seznam dot cz
GCC build triplet: x86_64-pc-linux-gnu i686-pc-linux-gnu
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30976
