14.2 Comparing newer versions with old ones

Q: I switched to v2 and my programs now run slower than when compiled with v1.x....

Q: I timed a test program and it seems that GCC 2.8.1 produces slower executables than GCC 2.7.2.1 was, which in turn was slower than DJGPP v1.x. Why are we giving up so much speed as we get newer versions?

Q: I installed Binutils 2.8.1, and my programs are now much slower than when they are linked with Binutils 2.7!

A: In general, newer versions of GCC generate tighter, faster code, than older versions. Comparison between different versions of GCC shows that they all optimize reasonably well, but it takes a different combination of the optimization-related options to achieve the greatest speed in each compiler version. The default optimization options can also change; for example, --force-mem is switched on by -O2 in 2.7.2.1; it wasn't before. GCC offers a plethora of optimization options which might make your code faster or slower (see the GCC docs for a complete list); the best way to find the correct combination for a given program is to profile and experiment. Here are some tips:

• Compile your code using the GCC switches -O2 -mpentium -fomit-frame-pointer -ffast-math. (For PGCC and GCC version 2.95 and later, use -O6 instead of -O2.)
• Profile your code and check which functions are "hot spots". Disassemble them, or compile with -S (see getting assembly listing), and examine the machine code.
• If the disassembly shows there aren't too many memory accesses inside the inner loops, try adding -fforce-addr option. This option helps a lot if a couple of pointers are used heavily within a single loop. If there are a lot of memory references, try adding -fno-force-mem, to prevent GCC from repeatedly copying variables from memory into registers.
• Sometimes, -fomit-frame-pointer might make things worse, since it uses stack-relative addresses which have longer encoding and could therefore overflow the CPU cache. So try with and without this switch.
• With newer versions of GCC, programs whose inner loops include many function calls, or which are deeply recursive, could benefit from using the -mpreferred-stack-boundary=2 compiler option. This causes the compiler to relax its stack-alignment requirements that need a lot of sub esp,xx instructions. The default stack alignment is 16 bytes, unless overridden by -mpreferred-stack-boundary. The argument to this option is the power of 2 used for alignment, so 2 means 4-byte alignment; if your code uses double and long double variables, an argument of 3 might be a better choice.
• Depending on the structure of your code, try different combinations of alignment for loops (-malign-loops), jumps (-malign-jumps), and function entry points (-malign-functions). Alignment changes can have especially profound effects when programs are run on AMD's K6 CPU, since these CPUs suffer significant slowdown for code aligned on 4-byte boundaries.
• Try adding -funroll-loops and -funroll-all-loops and profile the effect.
• Try compiling with -fno-strength-reduce. In some cases where GCC is in dire need of registers, this could be a substantial win, since strength reduction typically results in using additional registers to replace multiplication with addition.
• If different time-critical functions exhibit different behavior under some of the optimization options, try compiling them with the best combination for each one of them.

I'm told that the PGCC version of GCC has bugs in its optimizer which show when you use level 7 or higher. Until that is solved in some future version, you are advised to stick to -O6. Some programs actually run faster when compiled with -O2 or -O3, even when compiled with PGCC, so you might try that as well. Several users reported that PGCC v2.95.1 tends to crash a lot during compilation, especially with -O5, -O6 and -mpentium options. (In general, PGCC version 2.95 is deemed buggy; you are advised not to use it.)

Programs which manipulate multi-dimensional arrays inside their innermost loops can sometimes gain speed by switching from dynamically allocated arrays to static ones. This can speed up code because the size of a static array is known to GCC at compile time, which allows it to avoid dedicating a CPU register to computing offsets. This register is then available for general-purpose use.

Another problem that is related to C++ programs which manipulate arrays happens when you fail to qualify the methods used for array manipulation as inline. Each method or function that wasn't declared inline will not be inlined by GCC, and will incur an overhead of a function call at run time.

However, inlining only helps with small functions/methods; large inlined functions will overflow the CPU cache and typically slow down the code instead of speeding it up.

If your CPU is AMD's K6, try upgrading to GCC 2.96 or later and use the -mcpu=k6 switch. I'm told that K6-specific optimizations are much better in these versions of GCC.

A bug in the startup code distributed with DJGPP versions before v2.02 can also be a reason for slow-down. The problem is that the runtime stack of DJGPP programs was not guaranteed to be properly aligned. This usually only shows up on Windows (since CWSDPMI aligns the stack on its own), and even then only sometimes. But it has been reported that switching to Binutils 2.8.1 sometimes causes such slow-down, and switching to PGCC can reveal this problem as well. In some cases, restarting Windows would cause programs run at normal speed again. If you experience such problems too much, upgrade to v2.02.