When should add padding to struct in LLVM IR? And When shouldn't?

Clang has different behaviors in different cases, when add padding to struct, and what's the rule?

For below c++ code:

struct CT1 {
  char   c1 = 'c';
  double d1;
  char   c2;
};

struct CT2 {
  char   c1;
  double d1;
  char   c2;
};
struct CT3 {
  char   c1 = 'c';
  double d1;
};

int main() {
  CT1 ct1;
  CT2 ct2;
  CT3 ct3;
  return 0;
}

Clang will generate IR code:

%struct.CT1 = type <{ i8, [7 x i8], double, i8, [7 x i8] }>
%struct.CT2 = type { i8, double, i8 }
%struct.CT3 = type { i8, double }

I have several questions:

1. CT1 has padding, while CT2 has not. CT1 is only different with CT2 by having a initial value of c1. Why?
2. CT3 hasn't padding, which is only different with CT1 by reduced a property named c2. Why?
3. IR Code of CT2 doesn't add padding explicitly, but in log (using -Wpadded) it seems to have added padding already, is this true? And does it means that we can not add padding by hand?

warning: padding struct 'CT3' with 7 bytes to align 'd1' [-Wpadded]
  double d1;

4. How to add padding to struct by IRBuilder of LLVM?

I need to generate IR code by hand, so i need to know when should i add padding to struct.

Really thanks for any reply.