Getting GCC / Clang to use CMOV

Question

Getting GCC / Clang to use CMOV

I have a simple labeled union of values. Values can be int64_ts

or doubles

. I am doing the addition in these unions with the caveat that if both arguments represent values int64_t

, then the result must also matter int64_t

.

Here is the code:

#include<stdint.h>
union Value {
  int64_t a;
  double b;
};

enum Type { DOUBLE, LONG };

// Value + type.
struct TaggedValue {
  Type type;
  Value value;
};

void add(const TaggedValue& arg1, const TaggedValue& arg2, TaggedValue* out) {
  const Type type1 = arg1.type;
  const Type type2 = arg2.type;
  // If both args are longs then write a long to the output.
  if (type1 == LONG && type2 == LONG) {
    out->value.a = arg1.value.a + arg2.value.a;
    out->type = LONG;
  } else {
    // Convert argument to a double and add it.
    double op1 = type1 == LONG ? (double)arg1.value.a : arg1.value.b; // Why isn't CMOV used?
    double op2 = type2 == LONG ? (double)arg2.value.a : arg2.value.b; // Why isn't CMOV used? 
    out->value.b = op1 + op2;
    out->type = DOUBLE;
  }
}

The gcc output on -O2 is here: http://goo.gl/uTve18 Attached here if the link doesn't work.

add(TaggedValue const&, TaggedValue const&, TaggedValue*):
    cmp DWORD PTR [rdi], 1
    sete    al
    cmp DWORD PTR [rsi], 1
    sete    cl
    je  .L17
    test    al, al
    jne .L18
.L4:
    test    cl, cl
    movsd   xmm1, QWORD PTR [rdi+8]
    jne .L19
.L6:
    movsd   xmm0, QWORD PTR [rsi+8]
    mov DWORD PTR [rdx], 0
    addsd   xmm0, xmm1
    movsd   QWORD PTR [rdx+8], xmm0
    ret
.L17:
    test    al, al
    je  .L4
    mov rax, QWORD PTR [rdi+8]
    add rax, QWORD PTR [rsi+8]
    mov DWORD PTR [rdx], 1
    mov QWORD PTR [rdx+8], rax
    ret
.L18:
    cvtsi2sd    xmm1, QWORD PTR [rdi+8]
    jmp .L6
.L19:
    cvtsi2sd    xmm0, QWORD PTR [rsi+8]
    addsd   xmm0, xmm1
    mov DWORD PTR [rdx], 0
    movsd   QWORD PTR [rdx+8], xmm0
    ret

He created code with a lot of branches. I know the input is pretty random, meaning it has a random combination int64_t

and double

s. I would like to have at least a conversion to double done with the equivalent of an instruction CMOV

. Is there a way to persuade gcc to generate this code? Ideally, I would like to run some real data benchmark to see how code with a lot of branches works against one with fewer branches but more expensive instructions CMOV

. It may turn out that the default code generated by GCC works better, but I would like to confirm this. I could inline the assembly myself, but I would rather not.