https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114404
Bug ID: 114404
Summary: [11] GCC reorders stores when it probably shouldn't
Product: gcc
Version: 11.4.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: iii at linux dot ibm.com
Target Milestone: ---
Reproducible with gcc commit 1b5510a59163.
I'm writing this up as a result of the following linux kernel discussion:
https://lore.kernel.org/bpf/[email protected]/
https://lore.kernel.org/bpf/[email protected]/
In the following code:
extern const char bpf_plt[];
extern const char bpf_plt_ret[];
extern const char bpf_plt_target[];
static void bpf_jit_plt(void *plt, void *ret, void *target)
{
memcpy(plt, bpf_plt, BPF_PLT_SIZE);
*(void **)((char *)plt + (bpf_plt_ret - bpf_plt)) = ret;
*(void **)((char *)plt + (bpf_plt_target - bpf_plt)) = target ?: ret;
}
GCC 11's sched1 pass reorders memcpy() and assignments. In GCC 12 this
behavior is gone after
commit 2e96b5f14e4025691b57d2301d71aa6092ed44bc
Author: Aldy Hernandez <[email protected]>
Date: Tue Jun 15 12:32:51 2021 +0200
Backwards jump threader rewrite with ranger.
but this seems to be accidental. Internally, output_dependence() for the
respective mems returns false, because it believes that they are based on
different SYMBOL_REFs. This may be because on the C level we are not allowed
to subtract pointers to different objects.
However, a possible solution to this should be casting pointers to longs, since
C pointer subtraction rules would no longer apply, but in practice this does
nothing.
In the attached minimized preprocessed source with long casts we get:
stg %r3,232(%r2,%r15)
ltgr %r11,%r11
locgrne %r3,%r11
stg %r3,232(%r1,%r15)
la %r2,0(%r1,%r9)
la %r3,232(%r1,%r15)
mvc 232(16,%r15),0(%r5)
mvc 248(16,%r15),16(%r5)
lghi %r4,8
brasl %r14,s390_kernel_write@PLT
so the assignments are placed before the memcpy().