Changing source code at compilation time (using LLVM)

Question

#include <stdio.h>
#include <string.h>    
int foo(char* a)
    { char str[10];
      if (strlen(a)<10) 
     { 
       sprintf(str,"Yes");
       puts(str);
       return 0;
     }
     else
     {
       sprintf(str,"No");
       puts(str);
       return 1;
     }
    }

Now, lets say while writing a LLVM pass, I want to ensure that instead of calling sprintf, printf is called instead (with the same arguments). How could I go about doing that?

Answer 1

In a nutshell,

1. Go over all instructions in the function.
2. If the instruction is a CallInst, check if it's a call to sprintf (you can just check its name).
3. Create a new CallInst (via CallInst::Create) that calls printf instead of sprintf. I think the easiest way to get the Value of the printf declaration is to use one of the Module::getOrCreate methods. The type of printf should be the same as the one for sprintf minus the first parameter.
4. Set the operands of the new call instructions to be the same as for sprintf, minus the first parameter.
5. Replace the old call with the new call - replaceInstWithInst (in BasicBlockUtils.h) is probably the most convenient way.
6. Finally, you might want to track the usages of the old first parameter, remove them all and then remove it - so you'd get rid of that puts call.

Answer 2

#ifdef USE_BUFFER
#define my_printf(...) sprintf(buffer, __VA_ARGS__)
#else
#define my_printf(...) printf(__VA_ARGS__)
#endif

Now you can use for instance my_printf("My name is %s.", "Bozo"); and it will compile as if it was printf("My name is %s.", "Bozo") by default.

If you include #define USE_BUFFER before the headers, it will instead convert those lines to sprintf(buffer, "My name is %s.", "Bozo") at compile time. Of course the variable buffer must exist in the context.