在osx上使用clang编译后的大堆栈框架大小导致堆栈溢出

Question

我有两个类似的递归函数：

void dfs(const std::vector<std::vector<int64_t>>& tree, int64_t position) {
  for (int64_t to : tree[position]) {
    dfs(tree, to);
  }

  if (position % 2 == 1) {
    for (int64_t to : tree[position]) {
    }
  } else {
    for (int64_t to : tree[position]) {
    }
  }
}

没有循环范围的相同代码：

void dfs(const std::vector<std::vector<int64_t>>& tree, int64_t position) {
  for (size_t i = 0; i < tree[position].size(); ++i) {
    int64_t to = tree[position][i];
    dfs(tree, to);
  }

  if (position % 2 == 1) {
    for (size_t i = 0; i < tree[position].size(); ++i) {
      int64_t to = tree[position][i];
    }
  } else {
    for (size_t i = 0; i < tree[position].size(); ++i) {
      int64_t to = tree[position][i];
    }
  }
}

完整代码here。

不幸的是第一个是segfaulting ，但仅限于带有clang的osx（3.5,3.8,3.9）。我使用gdb检查了堆栈帧大小：

Stack level 1, frame at 0x7fff5fbfebe0:
rip = 0x100000a8c in dfs(std::__1::vector<std::__1::vector<long long, std::__1::allocator<long long> >, std::__1::allocator<std::__1::vector<long long, std::__1::allocator<long long> > > > const&, long long); saved rip = 0x100000a8c
called by frame at 0x7fff5fbfef00, caller of frame at 0x7fff5fbfe8c0
Arglist at 0x7fff5fbfebd0, args: 
Locals at 0x7fff5fbfebd0, Previous frame's sp is 0x7fff5fbfebe0
Saved registers:
rbp at 0x7fff5fbfebd0, rip at 0x7fff5fbfebd8

看起来第一个函数的dfs中的帧大小是：0x7fff5fbfef00 - 0x7fff5fbfebe0 = 0x320 = 800 为什么第一个代码使用接近1kb的堆栈帧？第二个代码仅使用：0x7fff5fbff240 - 0x7fff5fbff140 = 0x100 = 256字节。

Stack level 1, frame at 0x7fff5fbff140:
rip = 0x100000d4d in dfs(std::__1::vector<std::__1::vector<long long, std::__1::allocator<long long> >, std::__1::allocator<std::__1::vector<long long, std::__1::allocator<long long> > > > const&, long long); saved rip = 0x100000d4d
called by frame at 0x7fff5fbff240, caller of frame at 0x7fff5fbff040
Arglist at 0x7fff5fbff130, args: 
Locals at 0x7fff5fbff130, Previous frame's sp is 0x7fff5fbff140
Saved registers:
rbp at 0x7fff5fbff130, rip at 0x7fff5fbff138

PS：编译标志： clang ++ -O0 -std = c ++ 11 ./file.cpp

PSS：我试图尽可能地缩小代码。例如，如果我在其中一个分支中删除一个空的循环范围，那么一切都会恢复正常。

PSSS：如果我用g ++ - 5编译，dfs中的帧大小只有0x50 = 128字节。同样在Linux上，一切看起来都不错。