Code optimization; switch versus if's

Question

I have a question about whether to use 'case' or 'ifs' in a function that gets called quite a lot. Here's the following as it is now, in 'ifs'; the code is self-explanatory:

int identifyMsg(char* textbuff) {
if (!strcmp(textbuff,"text")) {
    return 1;
}
if (!strcmp(textbuff,"name")) {
    return 2;
}
if (!strcmp(textbuff,"list")) {
    return 3;
}
if (!strcmp(textbuff,"remv")) {
    return 4;
}
if (!strcmp(textbuff,"ipad")) {
    return 5;
}
if (!strcmp(textbuff,"iprm")) {
    return 6;
}
return 0;
}

My question is: Would a switch perform better? I know if using 'ifs', I can place the most likely options at the top.

Answer 1

You can't use switch statements for strings because they are pointers and are not evaluted at compile time. You are stuck with using a bunch of if statements.

However for the sake of performance, I believe switches perform better when there are more conditions to check but the difference will be so minute it wouldn't matter.

I've never tested this before though, but I've read about this kind of switch optimization:

switch (value) {
  case frequent_value1:
  case frequent_value2:
  case frequent_value3:
    break;

default:
  switch (value) {
     case infrequent_value1:
     case infrequent_value2:
     case infrequent_value3:
        break;
     }
}

Answer 2

You could use gperf to generate perfect hashes of the "verbs" you want to see. Then you could use a switch statement.

Or, you could do something like this:

switch (textbuff[0])
{
    case 'i':
    {
        switch (textbuff[1])
        {
            case 'p':
            {
                 switch (textbuff[2])
                 {
                     case 'a': /* something. */ break;
                     case 'r': /* something else. */ break;
                 }
            }
        }
    }

(You get the idea).

As yet another option (if all of your commands are 4 characters), turn them into a single 32-bit number and then switch on that:

int32_t mashed =
    textbuff[0] << 24 | 
    textbuff[1] << 16 |
    textbuff[2] << 8 |
    textbuff[3];

switch (mashed) { /* ... */ }

To be honest, though, unless the list of options is particularly large, or this function is being called an obscene number of times, it won't be worth it.

Remember: measure first; optimise later (only if necessary).

Answer 3

You could put all the values into a std::map.

class MessageMap
{ 
    std::map<std::string,int>    data;
    public:
        MessageMap()
        {
             data["text"]   = 1;
             data["name"]   = 2;
             data["list"]   = 3;
             data["remv"]   = 4;
             data["ipad"]   = 5;
             data["iprm"]   = 6;
        }
        int getMessageId(std::string const& index) cosnt
        {
            std::map<std::string,int>::const_iterator f;
            if ((f = data.find(index)) != data.end())
            {
                return f->second;
            }
            return 0;
        }
};
int identifyMsg(char* textbuff)
{
    static MessageMap   mssageMap;
    return messageMap.getMessageId(textbuff);
}

Answer 4

You could very well use "enum" for those strings. then use switch case statements.

Answer 5

Another alternative that I have come across recently which might or might not fit your liking is:

int identifyMsg(const char* textbuff) {
    static const struct { const char* str; int id; } pairs[] = {
        { "text", 1 },
        { "name", 2 },
        { "list", 3 },
        { "remv", 4 },
        { "ipad", 5 },
        { "iprm", 6 },
    };
    for (int i = 0; i < sizeof(pairs)/sizeof(pairs[0]); ++i) {
        if (!strcmp(textbuff, pairs[i].str))
            return pairs[i].id;
    }
    return 0;
}

Answer 6

There were two questions, as far as I understood. Optimization and if/switch.

First of all, code-optimization is a costly process. Optimize only those parts of code, which are bottle-necks by evident. I doubt it in this case. Looks, like you are dispatching a textual-id for making a decision what to do with a message. Where does the message come from? IPC, XML, File? What are you going to do with this message? How performant is the message-content processing-code? There should be places in code, which take more ressources than string comparison.

Have you tried out some performance analyzers like Purify, gperf, cachegrind?

Concerning if/switch: switch works only with integer-types. (char, short, int, long, enum)

Answer 7

Assuming it actually matters:

Because strcmp is slow, but an integer compare is fast: if all your commands are 4 characters long - which happens to fit in a 32bit integer - you can cast each string to a 32bit number, and then switch based on that.

Otherwise, there are two basic ways to quickly compare a string against a lot of candidate strings:

• Store the candidates in a hash table.

or

• Sort the candidates alphabetically sorted in an array. You can then perform a binary search on the array by using the result of strcmp to either find a match, or exclude half of the remaining candidates.

---

As a side note - Compilers, such as MSVC and GCC, have implemented an optimization on switches that tests the switch conditions using a binary search. So a switch statement with, say, 256 elements, will be optimized down to, at most, 8 compare operations.