[Bug c++/111577] New: -Os gives significantly bigger code than -O0
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111577
Bug ID: 111577
Summary: -Os gives significantly bigger code than -O0
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: socketpair at gmail dot com
Target Milestone: ---
Yes, I saw #35806, #41175 and others.
See https://godbolt.org/z/Pnh89Y3Yb
```
#include
using namespace std;
int main(int argc, char* argv[]) {
if (argv[0] == nullptr || argv[1] == nullptr) return 0;
string zxc(argv[0]);
string qwe(argv[1]);
string asd(argv[2]);
zxc = qwe + asd;
return zxc.size();
}
```
-Os -std=c++2b -march=skylake -m64 (615 bytes)
compare size with the smae, but with options:
-O0 -std=c++2b -march=skylake -m64 (409 bytes)
-O0 - is much LESS (!) in bytes. I think it's a bug.
[Bug tree-optimization/108418] gcc does not optimize trivial code
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108418 --- Comment #1 from Коренберг Марк --- Sorry, but such kind of code happens as a result of C-code automatic generation.
[Bug tree-optimization/108418] New: gcc does not optimize trivial code
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108418
Bug ID: 108418
Summary: gcc does not optimize trivial code
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: socketpair at gmail dot com
Target Milestone: ---
https://godbolt.org/z/s3j8jK6ca
```
#include
int firewall1(const uint8_t *restrict data) {
const uint8_t ip_proto = *data;
const uint16_t dst_port = *((const uint16_t *)data + 32);
const uint16_t qwe = *((const uint16_t *)data + 64);
if (ip_proto == 17 && dst_port == 17 && qwe == 42) return 1;
if (ip_proto == 17 && dst_port == 23 && qwe == 42) return 1;
if (ip_proto == 17 && dst_port == 45 && qwe == 42) return 1;
if (ip_proto == 17 && dst_port == 63 && qwe == 42) return 1;
if (ip_proto == 17 && dst_port == 0 && qwe == 42) return 1;
if (ip_proto == 17 && dst_port == 2 && qwe == 42) return 1;
if (ip_proto == 17 && dst_port == 3 && qwe == 42) return 1;
return 0;
}
int firewall2(const uint8_t *restrict data) {
const uint8_t ip_proto = *data;
const uint16_t dst_port = *((const uint16_t *)data + 32);
const uint16_t qwe = *((const uint16_t *)data + 64);
if (ip_proto == 17 && dst_port == 17) return 1;
if (ip_proto == 17 && dst_port == 23) return 1;
if (ip_proto == 17 && dst_port == 45) return 1;
if (ip_proto == 17 && dst_port == 63) return 1;
if (ip_proto == 17 && dst_port == 0) return 1;
if (ip_proto == 17 && dst_port == 2) return 1;
if (ip_proto == 17 && dst_port == 3) return 1;
return 0;
}
```
It can't understand common condition (ip_proto == 17 && qwe == 42).
But it can for simpler case in firewall2.
See godbolt assembler output.
[Bug tree-optimization/108215] New: Does not optimize trivial case with bit operations
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108215
Bug ID: 108215
Summary: Does not optimize trivial case with bit operations
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: socketpair at gmail dot com
Target Milestone: ---
https://godbolt.org/z/5e3eKqPqs
```C
#include
int firewall3(const uint8_t *restrict data) {
const uint32_t src = *((const uint32_t *)data);
if ((src & 0x) == 0x1122) return 1;
if ((src & 0xFF00) == 0x11223300) return 1;
return 0;
}
int firewall4(const uint8_t *restrict data) {
const uint32_t src = *((const uint32_t *)data);
if ((src & 0xFF00) == 0x11223300) return 1;
if ((src & 0x) == 0x1122) return 1;
return 0;
}
```
```
firewall3:
movl(%rdi), %eax
xorw%ax, %ax
cmpl$287440896, %eax
sete%al
movzbl %al, %eax
ret
firewall4:
movl(%rdi), %eax
movl$1, %edx
movl%eax, %ecx
xorb%cl, %cl
cmpl$287453952, %ecx
je .L3
xorw%ax, %ax
xorl%edx, %edx
cmpl$287440896, %eax
sete%dl
.L3:
movl%edx, %eax
ret
```
firewall3(): Excellent!
firewall4(): FAIL!
It's obvious that order of comparisons in this example does not matter. So I
think misoptimisation of firewall4() is a bug.
[Bug tree-optimization/107767] [13 Regression] switch to table conversion happening even though using btq is better
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107767 --- Comment #8 from Коренберг Марк --- Okay, but why switch-case is not handled using fast implementation using masks (when difference between smallest and biggest integer <=64 ? See the first function in my first message where it works as expected. Seems, the problem is not in converting to switch-case, but missing optimisation for switch-case case.
[Bug tree-optimization/107767] [13 Regression] switch to table conversion happening even though using btq is better
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107767 --- Comment #5 from Коренберг Марк --- Not only -s problem. I think -O3 in gcc 12.2 will run faster than -O3 in gcc 13 (for this case). this code should not be treated as if-else-if-else-if. gcc 12 does its job right.
[Bug regression/107767] GCC has some problems in optimizer of trivial case
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107767 --- Comment #3 from Коренберг Марк --- I forgot to add. gcc 12.2 - everything is OK. gcc 13 - with bug.
[Bug regression/107767] GCC has some problems in optimizer of trivial case
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107767 --- Comment #1 from Коренберг Марк --- See assembler output for firewall2(). It's not -Os optimized (compare to firewall(), which is ok) ``` firewall: movw64(%rdi), %ax cmpb$17, (%rdi) sete%cl leal-15(%rax), %edx testw $-9, %dx movb$1, %dl sete%sil cmpw$47, %ax ja .L2 movabsq $-180319906955279, %rdx btq %rax, %rdx setc%dl .L2: movl%edx, %eax xorl$1, %eax orl %esi, %eax andl%ecx, %eax movzbl %al, %eax ret firewall2: movw64(%rdi), %ax xorl%edx, %edx decl%eax cmpw$46, %ax ja .L5 movzwl %ax, %eax movsbl CSWTCH.2(%rax), %edx .L5: movl%edx, %eax ret CSWTCH.2: .byte 1 .byte 1 .byte 1 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 1 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 1 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 0 .byte 1 .byte 0 .byte 0 .byte 1 .byte 0 .byte 1 ```
[Bug regression/107767] New: GCC has some problems in optimizer of trivial case
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107767
Bug ID: 107767
Summary: GCC has some problems in optimizer of trivial case
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: regression
Assignee: unassigned at gcc dot gnu.org
Reporter: socketpair at gmail dot com
Target Milestone: ---
See https://godbolt.org/z/rTfTondfP
```
#include
int firewall(const uint8_t *restrict data) {
const uint8_t ip_proto = *data;
const uint16_t dst_port = *((const uint16_t *)data + 32);
if (ip_proto == 17 && dst_port == 15) return 1;
if (ip_proto == 17 && dst_port == 23) return 1;
if (ip_proto == 17 && dst_port == 47) return 1;
if (ip_proto == 17 && dst_port == 45) return 1;
if (ip_proto == 17 && dst_port == 42) return 1;
if (ip_proto == 17 && dst_port == 1) return 1;
if (ip_proto == 17 && dst_port == 2) return 1;
if (ip_proto == 17 && dst_port == 3) return 1;
return 0;
}
int firewall2(const uint8_t *restrict data) {
const uint16_t dst_port = *((const uint16_t *)data + 32);
if (dst_port == 15) return 1;
if (dst_port == 23) return 1;
if (dst_port == 47) return 1;
if (dst_port == 45) return 1;
if (dst_port == 42) return 1;
if (dst_port == 1) return 1;
if (dst_port == 2) return 1;
if (dst_port == 3) return 1;
return 0;
}
```
Compile with -Os.
Second function IS NOT minimal, obviously. It's a bug. GCC 12.2 does not have
it.
