https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72795
Bug ID: 72795
Summary: Missed optimization of external-linkage variables in
presence of barriers
Product: gcc
Version: 6.1.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: ahmad at a3f dot at
Target Milestone: ---
Godbolt link: https://godbolt.org/g/egb73f
int x;
static int y;
int f() { x = 1; asm volatile("":::"memory"); x = 2; return x; }
int g() { y = 1; asm volatile("":::"memory"); y = 2; return y; }
int h() { int z = 1; asm volatile("":::"memory"); z = 2; return z; }
on GCC 6.1 with -O2 for AMD64 generates following assembly:
f():
movl $1, x(%rip)
movl $2, x(%rip)
movl $2, %eax
ret
g():
movl $2, %eax
ret
h():
movl $2, %eax
ret
The question is: Was optimizing the non-volatile write in g() and h() the
intended behavior?
And if so, shouldn't same optimization be applied when the variable is of
external linkage (f)?
Replacing the compiler barrier with a full memory barrier gives same behavior.
I also tried same with clang and icc. Clang does as GCC does, icc doesn't
optimize g(), but optimizes h().
