https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119703
Bug ID: 119703
Summary: x86: spurious branches for inlined memset in ranges
(40; 64) when requesting unrolled loops without simd
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: mjguzik at gmail dot com
Target Milestone: ---
13.3.0 runs into it, but I also tested on godbolt which claims to have 15.0.1:
gcc
(Compiler-Explorer-Build-gcc-ca4e6e6317ae0ceada8c46ef5db5ece165a6d1c4-binutils-2.42)
15.0.1 20250409 (experimental)
... and got the same result.
I have not verified memcpy, I suspect it might suffer the same problem.
src:
void zero(char *buf)
{
__builtin_memset(buf, 0, SIZE);
}
compiled like so:
cc -O2 -DSIZE=48 -mno-sse
-mmemset-strategy=unrolled_loop:256:noalign,libcall:-1:noalign -c zero.c
objdump says:
0000000000000000 <zero>:
0: f3 0f 1e fa endbr64
4: 31 d2 xor %edx,%edx
6: 31 c0 xor %eax,%eax
8: 89 c1 mov %eax,%ecx
a: 48 89 14 0f mov %rdx,(%rdi,%rcx,1)
e: 48 89 54 0f 08 mov %rdx,0x8(%rdi,%rcx,1)
13: 48 89 54 0f 10 mov %rdx,0x10(%rdi,%rcx,1)
18: 48 89 54 0f 18 mov %rdx,0x18(%rdi,%rcx,1)
1d: 83 c0 20 add $0x20,%eax
20: 72 e6 jb 8 <zero+0x8>
22: 48 01 c7 add %rax,%rdi
25: 48 c7 07 00 00 00 00 movq $0x0,(%rdi)
2c: 48 c7 47 08 00 00 00 movq $0x0,0x8(%rdi)
33: 00
34: c3 ret
As you can see it emits body for a 32-byte loop, bumps eax and branches on it.
but it knows eax is 0, in fact it zeroed it itself. Then it fills up the
trailing 16 bytes.
I verified this does not happen with 40:
0000000000000000 <zero>:
0: f3 0f 1e fa endbr64
4: 48 c7 07 00 00 00 00 movq $0x0,(%rdi)
b: 48 c7 47 08 00 00 00 movq $0x0,0x8(%rdi)
12: 00
13: 48 c7 47 10 00 00 00 movq $0x0,0x10(%rdi)
1a: 00
1b: 48 c7 47 18 00 00 00 movq $0x0,0x18(%rdi)
22: 00
23: 48 c7 47 20 00 00 00 movq $0x0,0x20(%rdi)
2a: 00
2b: c3 ret
this looks fine.
Note going past 40 is the magic threshold where it starts looking at rep unless
told otherwise.
Zeroing 41 bytes is also bad:
0000000000000000 <zero>:
0: f3 0f 1e fa endbr64
4: 31 d2 xor %edx,%edx
6: 31 c0 xor %eax,%eax
8: 89 c1 mov %eax,%ecx
a: 48 89 14 0f mov %rdx,(%rdi,%rcx,1)
e: 48 89 54 0f 08 mov %rdx,0x8(%rdi,%rcx,1)
13: 48 89 54 0f 10 mov %rdx,0x10(%rdi,%rcx,1)
18: 48 89 54 0f 18 mov %rdx,0x18(%rdi,%rcx,1)
1d: 83 c0 20 add $0x20,%eax
20: 72 e6 jb 8 <zero+0x8>
22: 48 01 c7 add %rax,%rdi
25: 48 c7 07 00 00 00 00 movq $0x0,(%rdi)
2c: c6 47 08 00 movb $0x0,0x8(%rdi)
30: c3 ret
