9/18    I have an embedded C++ application that passes around a lot of
        68-byte structs-- the struct is a wrapper around a binary message,
        and just about every function call includes one of them (usually
        passed by copy).  Today I changed the struct into a class which
        contains an empty Constructor/Destructor, and Configure() method
        which initializes every field.  Basically nothing else was changed:
        the 68-byte struct became a 72-byte object.  No inheritance, no
        virtual functions, everything is allocated on the stack.  My test
        suite is taking 300% longer to run, even with -O3.  Any ideas
        about what might be causing this?  I'm going to look into gprof,
        but if there are any hints and tips I'd like to hear them.  Thanks.
        \_ WAG: From your description, I gather you changed the definition
           from: struct Foo { ... }; to: class Foo { ... };
           Try changing "class" back to "struct" and see if that changes
           anything.
           \_ It does-- it cuts the running time by 2/3.  I've got 2
              parallel (except for the class stuff) directories and I'm
              running the tests side by side.
              \_ Well, are you passing these structs by value?  Are you
                 constructing and deconstructing tons and tons of classes
                 now?  Calling all those empty constructors and deconstructors
                 can get quite expensive.  Also do you have RTTI turned on?
                 That could explain the extra bytes,
                 \_ I don't have RTTI on explicitly (how would I check?),
                    and I'm not using any template stuff.  The compiler is
                    gcc 3.2.  Is there a way to optimize the (de|con)structor
                    calls to nothing?  I had figured that -O2 would take
                    care of that for me.  Also, I think that extra 4 bytes
                    is just a pointer to the dispatch table, which is
                    completely expected. --op
                    \_ If you have no virtual functions you don't have a
                       dispatch table, and you shouldn't have a pointer.
                       Trust me when your basic math objects are classes
                       and you have umpty millions of them, doubling the
                       size for a dispatch pointer would be really annoying.
                       \_ Aaah!  You're totally right.  Out of habit I had
                          put "virtual ~Foo() { };".  I removed the virtual
                          and the code is now about 5-10% *faster* than the
                          struct version.  Thank you thank you thank you.
              \_ Two more suggestions: Did you define Configure() in the
                 header file? If so, move it to the implementation file
                 and see if that makes a difference. Also, try compiling
                 with -Os, which optimizes for size (at least on my
                 gcc).          - struct guy
                 \_ Configure is defined in the .cc file; I'll give -Os
                    a shot now.
                    \_ Try swapping the the definition of the ctor/dtor
                       and Configure from .cc to .h or vice-versa.
        \_ Why did you change working code in the first place?  That's where
           your real problem is.
           \_ Obviously so he could put C/C++ on his resume!