void foo (int& sz, int n)
{
for (++n; --n;)
sz += 5;
}
Produces (-Os):
_Z3fooRii:
incl %esi
jmp .L2
.L3:
addl $5, (%rdi)
.L2:
decl %esi
jne .L3
ret
gcc thinks that every little change to sz needs to be stored in memory, causing
the above inefficiency of emitting a loop instead of a multiplication. It would
be desirable to treat sz as a local variable and optimize as if:
void bar (int& sz, int n)
{
int tmp = sz;
for (++n; --n;)
tmp += 5;
sz = tmp;
}
Which compiles to:
_Z3barRii:
leal (%rsi,%rsi,4), %esi
addl %esi, (%rdi)
ret
much better code. If somebody really does need each variable access to be a
memory access, that variable should be declared volatile.
If you don't want to fix this, at least provide a variable attribute to tell
the compiler that it's ok to omit stores.
--
Summary: Non-volatile variables don't need to be constantly
modified
Product: gcc
Version: 4.5.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: msharov at users dot sourceforge dot net
GCC host triplet: x86_64-gnu-linux
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45243
