Fix atomic alignment

[372046933]

Description

Below is the smallest reproducible code. the __attribute__((packed)) on ncclGinProxyGfd_t makes the alignment of ncclGinProxyGfd_t equals to 1. The packed struct here is not needed for two reasons.

1. We will do atomic operations on the address later. packed breaks the alignment requirement of atomics
2. ncclGinProxyQword_t is already packed and size of it is 8 bytes. So size of ncclGinProxyGfd_t is 64 bytes. and packed is redundant here.

#include <atomic>

typedef union {
  uint64_t raw;
  struct {
    uint64_t v : 1;
    uint64_t resv : 63;
  } __attribute__((packed)) flag;
} ncclGinProxyQword_t;


static_assert(sizeof(ncclGinProxyQword_t) == sizeof(uint64_t),
              "sizeof(ncclGinProxyQword_t) != sizeof(uint64_t)");

typedef struct __attribute__((packed)) {
  ncclGinProxyQword_t qword[8];
} ncclGinProxyGfd_t;

static_assert(alignof(ncclGinProxyGfd_t) == 1,
              "Incorrect packed implementation");

void f( ncclGinProxyGfd_t *q) {
  ncclGinProxyQword_t qword;
  __atomic_load(&q[0].qword[0].raw, &qword.raw, __ATOMIC_RELAXED);
}

gcc does not give any warning. But clang does.

test_atomic.cc:24:3: warning: misaligned atomic operation may incur significant performance penalty; the expected alignment (8 bytes) exceeds the actual alignment (1 bytes) [-Watomic-alignment]
   24 |   __atomic_load(&q[0].qword[0].raw, &qword.raw, __ATOMIC_RELAXED);
      |   ^
1 warning generated.

Related Issues

If we are going to support clang, this packed must be removed, otherwise libatomic needs to be linked.
#1744

Changes & Impact

This PR does not affect GCC. For clang, the generated code does not need link to libatomic

Performance Impact

No difference for GCC. Significant performance improvement on clang since libatomic call is eliminated

[372046933]

Or

typedef struct alignas(8) __attribute__((packed)) {
  ncclGinProxyQword_t qword[8];
} ncclGinProxyGfd_t;

IMHO, removing these attribute makes it simple