https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123765
Bug ID: 123765
Summary: bpf: incorrect code generation for overlapping
load/store
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: kristerw at gcc dot gnu.org
Blocks: 118443
Target Milestone: ---
Target: bpf-unknown-none
The following function is miscompiled with -fno-strict-aliasing -O1 or higher:
typedef int vec __attribute__((vector_size(4*sizeof(int))));
typedef unsigned uvec __attribute__((vector_size(4*sizeof(int))));
void h(vec*p,uvec*q){
vec a = *p;
*q = (uvec)a;
}
Compiling this as
bpf-unknown-none-gcc -O1 -fno-strict-aliasing -mlittle-endian -mcpu=v4 -S
bug.c
produces the following assembly:
h:
r0 = *(u64 *) (r1+0)
*(u64 *) (r2+0) = r0
r1 = *(u64 *) (r1+8)
*(u64 *) (r2+8) = r1
exit
The problem is that *p and *q are only guaranteed to be 64-bit aligned
(according to the information in the GIMPLE IR), so overlap is possible. In
that case, the first store may overwrite bytes that have not yet been loaded.
Therefore, both loads must be executed before the first store, such as:
h:
r0 = *(u64 *) (r1+0)
r1 = *(u64 *) (r1+8)
*(u64 *) (r2+0) = r0
*(u64 *) (r2+8) = r1
exit
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118443
[Bug 118443] [Meta bug] Bugs triggered by and blocking more smtgcc testing
