https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85330
Bug ID: 85330
Summary: -Os generates bigger code than -O2 due to disabled
strlen pass
Product: gcc
Version: 8.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: msebor at gcc dot gnu.org
Target Milestone: ---
When the test case below is compiled with -O2, GCC emits object code that's
optimally efficient both in space and runtime. But when the same test case is
compiled with -Os it results in much bigger object code that will run far
slower than the -O2 equivalent. That's because the strlen optimization doesn't
run with -Os.
At the same time, it's possible to create test cases involving string functions
where the opposite is true (often involving long strings). I wonder if
enabling the strlen optimization to a limited extent at -Os (to at least track
string lengths without other transformations) would give closer to optimal
results.
$ cat z.c && gcc -O2 -fdump-tree-optimized=/dev/stdout -c z.c && objdump -d z.o
&& gcc -Os -fdump-tree-optimized=/dev/stdout -c z.c && objdump -d z.o
#define S "012345678"
int f (void)
{
char a[sizeof S];
__builtin_strcpy (a, S);
return __builtin_strlen (a);
}
;; Function f (f, funcdef_no=0, decl_uid=1958, cgraph_uid=0, symbol_order=0)
f ()
{
<bb 2> [local count: 1073741825]:
return 9;
}
z.o: file format elf64-x86-64
Disassembly of section .text:
0000000000000000 <f>:
0: b8 09 00 00 00 mov $0x9,%eax
5: c3 retq
;; Function f (f, funcdef_no=0, decl_uid=1958, cgraph_uid=0, symbol_order=0)
f ()
{
char a[10];
long unsigned int _1;
int _4;
<bb 2> [local count: 1073741825]:
__builtin_strcpy (&a, "012345678");
_1 = __builtin_strlen (&a);
_4 = (int) _1;
a ={v} {CLOBBER};
return _4;
}
z.o: file format elf64-x86-64
Disassembly of section .text:
0000000000000000 <f>:
0: 48 83 ec 18 sub $0x18,%rsp
4: be 00 00 00 00 mov $0x0,%esi
9: 48 8d 54 24 06 lea 0x6(%rsp),%rdx
e: 48 89 d7 mov %rdx,%rdi
11: e8 00 00 00 00 callq 16 <f+0x16>
16: 48 83 c9 ff or $0xffffffffffffffff,%rcx
1a: 48 89 c2 mov %rax,%rdx
1d: 31 c0 xor %eax,%eax
1f: 48 89 d7 mov %rdx,%rdi
22: f2 ae repnz scas %es:(%rdi),%al
24: 48 83 c4 18 add $0x18,%rsp
28: 48 89 c8 mov %rcx,%rax
2b: 48 f7 d0 not %rax
2e: 48 ff c8 dec %rax
31: c3 retq
