[Bug target/123032] New: [16 Regression] wrong code on libsodium-1.0.20 since r16-5780-g65a3849eb46df2 (possibly fixed by r16-5849-g4a1247a3b070cd)

[slyfox at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123032

            Bug ID: 123032
           Summary: [16 Regression] wrong code on libsodium-1.0.20 since
                    r16-5780-g65a3849eb46df2 (possibly fixed by
                    r16-5849-g4a1247a3b070cd)
           Product: gcc
           Version: 16.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: slyfox at gcc dot gnu.org
  Target Milestone: ---

A few days ago I noticed gcc-master to miscompile libsodium-1.0.20 on
i686-linux. By the time I extracted the smaller example it does not reproduce
on master any more after r16-5849-g4a1247a3b070cd.

In IRC I asked if the fix could be related, Andrew suspects it might not:

    pinskia | because there is no early break in the memmove like code

Filing the bug to confirm the change is intentional and not masking another
bug.

Thus, on r16-5780-g65a3849eb46df2 minimal reproducer fails as:

// $ cat a.c
typedef unsigned char u8;
typedef unsigned int u32;

typedef struct {
    u32 pad; // important
    u8  buf[64];
} s_t;

__attribute__((noipa))
void h_bad(s_t *state, const u8 *in, u32 inlen) {
    for (u32 i = 0; i < inlen; i++) {
        state->buf[i] = in[i];
    }
}

__attribute__((noipa))
void h_good(s_t *state, const u8 *in, u32 inlen) {
    for (u32 i = 0; i < inlen; i++) {
        asm volatile("" ::: "memory"); // break vectorizer
        state->buf[i] = in[i];
    }
}

void h(s_t *state, const u8 *in, u32 inlen) {
    s_t s1 = {};
    s_t s2 = {};
    h_good(&s1, in, inlen);
    h_bad(&s2, in, inlen);
    if (__builtin_memcmp(&s1, &s2, sizeof(s1)) != 0) {
        __builtin_trap();
    }
}

int main(void) {
    s_t s = {};
    const u8 in[] =
        "0123456789" /* 0 */
        " 0123456789"
        " 0123456789"
        " 0123456789"
        " 0123456789"
    ;
    for (u32 o = 0; o < 10; o++) {
        __builtin_memset(&s, 0, sizeof(s));
        h(&s, &in[o], 33);
    }
}

$ gcc/xgcc -Bgcc a.c -o a -O3 && ./a
Illegal instruction        (core dumped) ./a

$ gcc/xgcc -Bgcc -v
Reading specs from gcc/specs
COLLECT_GCC=gcc/xgcc
COLLECT_LTO_WRAPPER=gcc/lto-wrapper
Target: i686-unknown-linux-gnu
Configured with: /home/slyfox/dev/git/gcc/configure --disable-multilib
--disable-bootstrap --disable-lto --disable-libsanitizer --enable-languages=c
CFLAGS='-O1 -g0' CXXFLAGS='-O1 -g0' LDFLAGS='-O1 -g0'
--build=i686-unknown-linux-gnu --host=i686-unknown-linux-gnu
Thread model: posix
Supported LTO compression algorithms: zlib
gcc version 16.0.0 20251203 (experimental) (GCC)