[TBAA] Miss optimization for check redundancy load

[hstk30-hw]

https://godbolt.org/z/b6vfG688K

typedef long int intptr_t;
typedef unsigned char uint8_t;
typedef unsigned short int uint16_t;
typedef unsigned int uint32_t;
typedef unsigned long int uint64_t;

typedef struct {
    union {
        uint8_t u6_ucaddr[16];
        uint16_t u6_usaddr[8];
        uint32_t u6_uladdr[4];
        uint64_t u6_ulladdr[2];
    } u6_addr;
} In6Addr;

struct policy_match_cond {
    In6Addr src_ipv6;
    In6Addr dst_ipv6;
};

typedef struct mbuf {
    void *userdata;
} __attribute__((__aligned__(64))) Mbuf;

typedef struct {
    In6Addr src_ip;
    In6Addr dst_ip;
} __attribute__((__aligned__(64))) MbufContext;

void test(Mbuf *mbuf, struct policy_match_cond *condition)
{
(condition->dst_ipv6.u6_addr.u6_uladdr)[0] = ((((((((MbufContext *)((mbuf))->userdata))->dst_ip).u6_addr.u6_uladdr)[0])&0xFF000000) >> 24);
(condition->dst_ipv6.u6_addr.u6_uladdr)[1] = ((((((((MbufContext *)((mbuf))->userdata))->dst_ip).u6_addr.u6_uladdr)[1])&0xFF000000) >> 24);
}

Clang is one more load than GCC.

The primary cause maybe clang not handle union's TBAA, like issue #146652.

But in this case, we might be able to delete the second load in InstCombine/CSE/GVN step?

[hstk30-hw]

GVN say:

GVN iteration: 0
GVN: load ptr %0 has unknown dependence
GVN: load i32 %1 has unknown dependence
GVN: load ptr %2 is clobbered by   store i32 %shr, ptr %dst_ipv6, align 8, !tbaa !11
GVN: load i32 %3 is clobbered by   store i32 %shr, ptr %dst_ipv6, align 8, !tbaa !11

Maybe still need union TBAA.