On Thu, Nov 09, 2017 at 04:02:53PM -0500, Wei Wei wrote:
> Hi all,
>
> I get a compile time error after setting -Og when compiling for the latest
> GitHub version.
> I am using `make defconfig’ to get the default x86_64 config. But previously
> I did this in v4.4,
> it's fine.
; cat >a.c <<'EOF'
extern void __attribute((error("1"))) f1(void);
extern void __attribute((error("2"))) f2(void);
static inline __attribute__((always_inline)) void bar(const void *addr)
{
int sz = __builtin_object_size(addr, 0);
if (__builtin_expect(sz >= 0, 0))
f1();
if (sz >= 0)
f2();
}
void foo(int *in)
{
bar(in);
}
EOF
; gcc -O2 -c a.c
; gcc -Og -c a.c
In function ‘bar’,
inlined from ‘foo’ at a.c:15:2:
a.c:8:3: error: call to ‘f1’ declared with attribute error: 1
f1();
^
;
Note that the call of f2() _was_ eliminated. Wrap the condition into
__builtin_expect() and with -Og it doesn't get eliminated until too
late. It's really brittle and dependent not just upon the _result_
of optimizations (get rid of __attribute((error())) in those and
you'll see that with -Og it compiles into
.file "a.c"
.text
.globl foo
.type foo, @function
foo:
.LFB1:
.cfi_startproc
rep ret
.cfi_endproc
.LFE1:
.size foo, .-foo
getting rid of both calls, as it ought to); it depends upon the moment
when dead code elimination happens. -Og leaves the sucker around for
too long. If you replace that __builtin_object_size() with -1 (which
is what it evaluates to), you'll get elimination happening early enough
even with -Og; ditto for getting rid of inlining...
