Hard-coded hardware addresses in C/C++

I read a blog post “reinterpret_cast vs. constant expression” that discusses how to get rid of C-style casts for code such as

#define FOO ((struct S*)0xdff000)

But there is no need to have hard-coded addresses in the code – it is better to declare a normal structure

extern struct S hw_s;

and tell the linker to place it at address 0xdff000 using an assembly file containing the lines

.global hw_s
hw_s = 0xdff000

FOO can now be defined without a cast

#define FOO &hw_s

although it is probably better to use hw_s directly...

It is good to get rid of hard-coded addresses in C/C++ code even if you do not care about ugly casts. One reason is that the compiler cannot know which objects the hard-coded addresses point to, which restricts the data flow analysis. One other reason is that hard-coded addresses interact badly with instruction selection in the backend. This is especially true for code accessing hardware registers that expand to assignments of the form

*(volatile int *)(0xdff008) = 0;
*(volatile int *)(0xdff010) = 10;

The best way of generating the code depends on the CPU architecture, but it usually involves loading a base address into a register and storing using a “base + offset” addressing mode, so the compiler needs to split and re-combine the addresses (which is complicated as there are often restrictions on which offsets are valid, the cost of the base depends on the value, etc.). The ARM backend is good at this, but I have seen many cases where much slower and larger code than necessary is generated for more obscure architectures. For example, GCC 7.1 for RISC-V compiles

void foo(void)
{
  *(volatile int *)(0xfffff00023400008) = 0;
  *(volatile int *)(0xfffff00023400010) = 10;
}

to

foo:
 lui a5,%hi(.LC0)
 ld a5,%lo(.LC0)(a5)
 li a4,10
 sw zero,0(a5)
 lui a5,%hi(.LC1)
 ld a5,%lo(.LC1)(a5)
 sw a4,0(a5)
 ret
.LC0:
 .dword -17591594647544
.LC1:
 .dword -17591594647536

instead of the smaller and faster

foo:
 lui a5,%hi(.LC0)
 ld a5,%lo(.LC0)(a5)
 li a4,10
 sw zero,8(a5)
 sw a4,16(a5)
 ret
.LC0:
 .dword -17591594647552

you get by writing through a normal structure.

A look at range-v3 code generation

I recently saw a Stack Overflow post that compared the speed of std::find_if on a vector vec

auto accumulated_length = 0L;
auto found = std::find_if(vec.begin(), vec.end(),
                          [&](auto const &val) {
                              accumulated_length += val;
                              return to_find < accumulated_length;
                          });
auto const found_index = std::distance(vec.begin(), found);

and the equivalent code using the range-v3 library

auto const found_index = ranges::distance(vec
    | ranges::view::transform(ranges::convert_to<long>{})
    | ranges::view::partial_sum()
    | ranges::view::take_while([=](auto const i) {
          return !(to_find < i);
      }));

Measuring the performance on an Intel Broadwell CPU using the Google benchmark library and this code compiled with the options

-O3 -march=native -std=c++14 -DNDEBUG

gives me the result

Benchmark                Time           CPU Iterations
------------------------------------------------------
BM_std/1024            311 ns        311 ns    2248354
BM_range/1024         2102 ns       2102 ns     332711

for gcc 7.1.0 and

BM_std/1024            317 ns        317 ns    2208547
BM_range/1024          809 ns        809 ns     864328

for clang 4.0.0. There are two obvious questions

• Why is range-v3 slower than the STL?
• Why is the difference so much bigger for GCC than for LLVM?

I also wanted to see if the STL added overhead, so I tried a simple C-style for-loop

long i, acc = 0;
for (i = 0; i < len; i++) {
    acc += p[i];
    if (to_find < acc)
        break;
}
found_index = i;

This runs in 439 ns – 40% slower than the STL version! – which adds the question

• Why is the for-loop slower than the STL version?

Why is the for-loop slower?

GCC is generating the obvious assembly for the for-loop

.L4:
        movslq  (%r8,%rax,4), %rcx
        addq    %rcx, %rdx
        cmpq    %rsi, %rdx
        jg      .L7
.L3:
        addq    $1, %rax
        cmpq    %rdi, %rax
        jl      .L4
.L7:
        ...

I had expected the compiler to generate similar code for std::find_if, and that is what happens if it is used with an input iterator, but libstdc++ has an overload for random-access iterators which partially unrolls the loop

template<typename _RandomAccessIterator, typename _Predicate>
  _RandomAccessIterator
  __find_if(_RandomAccessIterator __first, _RandomAccessIterator __last,
            _Predicate __pred, random_access_iterator_tag)
  {
    typename iterator_traits<_RandomAccessIterator>::difference_type
      __trip_count = (__last - __first) >> 2;

    for (; __trip_count > 0; --__trip_count)
      {
        if (__pred(__first))
          return __first;
        ++__first;

        if (__pred(__first))
          return __first;
        ++__first;

        if (__pred(__first))
          return __first;
        ++__first;

        if (__pred(__first))
          return __first;
        ++__first;
      }

    switch (__last - __first)
      {
      case 3:
        if (__pred(__first))
          return __first;
        ++__first;
      case 2:
        if (__pred(__first))
          return __first;
        ++__first;
      case 1:
        if (__pred(__first))
          return __first;
        ++__first;
      case 0:
      default:
        return __last;
      }
  }

This partial unrolling gets rid of a large fraction of the comparisons and branches, which makes a big difference for this kind of micro-benchmark.

Why does GCC generate slow code for range-v3?

The range-v3 code generated by GCC have a few objects placed on the stack which adds some (useless) memory operations. The reason they are not optimized has to do with how GCC are optimizing structures and the order the optimization passes are being run.

The GCC “Scalar Replacement of Aggregates” (SRA) optimization pass splits structures into their elements. That is,

struct S {
  int a, b, c;
};

struct S s;

s.a = s.b = s.c = 0;
...

is transformed to the equivalent of

int a, b, c;

a = b = c = 0;
...

and the variables are then optimized and placed in registers in the same way as normal non-structure variables.

The compiler cannot split structures that have their address taken as it would then need to do expensive pointer tracking to find how each element is used, so such structures are kept on the stack. The GCC SRA pass is conservative and does not split a structure if any part of it has been captured by a pointer, such as

struct S s;

s.a = s.b = s.c = 0;
int *p = &s.a;
...

that could be split into

int a, b, c;

a = b = c = 0;
int *p = &a;
...

but that is not done by GCC.

It is usually not a problem that address-taking limits SRA as optimization passes such as constant propagation eliminates use of pointers when they are only used locally in a function, so code of the form

struct S s;

int *p = &s.a;
...
*p = 0;

is transformed to

struct S s;

...
s.a = 0;

which can then be optimized by SRA. But this requires that all paths to the use of p pass through the same initialization and that the compiler can see that they pass through the same initialization – we cannot easily eliminate the pointers for code such as

struct S s;
int *p;

if (cond)
    p = &s.a;
...
if (cond)
    *p = 0;

that need the compiler to track values to see that all executions of *p initializes p to &s.a.

And that is how the range-v3 code looks like after templates has been expanded and all functions inlined – the code does different initializations depending on if the range is empty or not and ends up with code segments of the form

if (begin != end) {
    // Initialize some variables
}

...

if (begin != end) {
    // Use the variables
}

I have a hard time trying to follow exactly what range-v3 is trying to do – the code expands to more than 700 functions, so I have only looked at the compiler’s IR after inlining and I do not know exactly how it look in the C++ source code – but the result is that the compiler fails to propagate some addresses due to this issue and three objects (one struct take_while_view and two struct basic_iterator) are still placed on the stack when the last SRA pass has been run.

GCC do eventually manage to simplify the code enough that SRA could eliminate all structures, but that is later in the optimization pipeline, after the last SRA pass has been run. I tested to add an extra late SRA pass – this eliminates the memory operations, and the function runs in 709 ns. Much better, but still only half the speed of the STL version.

Why is range-v3 slower than the STL?

Both GCC and LLVM generate the range-v3 code to something of the form

static long foo(const int *begin, const int *end, long to_find)
{
    long result = 0;
    const int *p = begin;
    if (begin != end) {
        result = *begin;
        while (1) {
            if (p == end)
                break;
            if (to_find < result)
                break;

            p++;
            if (p != end)
                result += *p;
        }
    }
    return p - begin;
}

that does one extra comparison in the loop body compared to the for-loop version. This kind of code is supposed to be simplified by the loop optimizers, but they are running relatively early in the optimization pipeline (partly so that later optimizations may take advantage of the improved loop structure, and partly as many optimizations makes life harder for the loop optimizer) so they are limited by the same issues mentioned in the previous section – that is, I assume the redundant comparison would be eliminated if the range-v3 library improved its handling of empty ranges etc.

-fipa-pta

My previous blog post had a minimal description of -fipa-pta and I have received several questions about what it actually do. This blog post will try to give some more details...

Points-to analysis

Many optimizations need to know if two operations may access the same memory address. For example, the if-statement in

i = 5;
*p = -1;
if (i < 0)
  do_something();

can be optimized away if *p cannot modify i.

GCC tracks what the pointers may point to using the general ideas from the paper “Efficient Field-sensitive pointer analysis for C”. I will not describe the details – the first few pages of the paper do it better than I can do here – but the principle is that each pointer is represented by a set of locations it may point to, the compiler is generating set constraints representing each statement in the program, and then solving those constraints to get the actual set of locations the pointer may point to.

But this process is expensive, so GCC is normally doing this one function at a time and assumes that called functions may access any memory visible to them.

-fipa-pta

The -fipa-pta optimization takes the bodies of the called functions into account when doing the analysis, so compiling

void __attribute__((noinline))
bar(int *x, int *y)
{
  *x = *y;
}

int foo(void)
{
  int a, b = 5;
  bar(&a, &b);
  return b + 10;
}

with -fipa-pta makes the compiler see that bar does not modify b, and the compiler optimizes foo by changing b+10 to 15

int foo(void)
{
  int a, b = 5;
  bar(&a, &b);
  return 15;
}

A more relevant example is the “slow” code from the “Integer division is slow” blog post

std::random_device entropySource;
std::mt19937 randGenerator(entropySource());
std::uniform_int_distribution<int> theIntDist(0, 99);

for (int i = 0; i < 1000000000; i++) {
  volatile auto r = theIntDist(randGenerator);
}

Compiling this with -fipa-pta makes the compiler see that theIntDist is not modified within the loop, and the inlined code can thus be constant-folded in the same way as the “fast” version – with the result that it runs four times faster.

Interprocedural optimization in GCC

Compilers can do a better job optimizing a function if they can use knowledge of other functions. The obvious case is inlining, but there are many more cases. This post lists the interprocedural optimizations implemented in GCC 7.

Many of the optimizations are only relevant for large functions (small functions are inlined into the caller!) or for helping other optimization passes. This makes it hard to give relevant examples, so the examples in this post are just illustrating the principles.

Parameter passing

Parameter passing for functions where GCC can see all callers (such as functions that are local to a translating unit, or when the whole program is compiled using link-time optimization) is optimized as

• Unused parameters are removed.
• Parameters passed  by reference may be changed to be passed by value. For example,

  static int foo(int *m)
  {
    return *m + 1;
  }
  
  int bar(void)
  {
    int i = 1;
    return foo(&i);
  }
  

  is changed to

  static int foo(int m)
  {
    return m + 1;
  }
  
  int bar(void)
  {
    int i = 1;
    return foo(i);
  }
  

  which makes it much easier for other optimization passes to reason about the variables.
• A structure may be split into its elements. For example,

  struct bovid
  {
    float red;
    int green;
    void *blue;
  };
  
  static void ox(struct bovid *cow)
  {
    cow->red = cow->red + cow->green;
  }
  
  int main(void)
  {
    struct bovid cow;
  
    cow.red = 7.4;
    cow.green = 6;
    cow.blue = &cow;
  
    ox(&cow);
  
    return 0;
  }
  

  is changed to

  struct bovid
  {
    float red;
    int green;
    void *blue;
  };
  
  static void ox(float *t1, int t2)
  {
    *t1 = *t1 + t2;
  }
  
  int main(void)
  {
    struct bovid cow;
  
    cow.red = 7.4;
    cow.green = 6;
    cow.blue = &cow;
  
    ox(&cow.red, cow.green);
  
    return 0;
  }

These optimizations are enabled by -fipa-sra, which is enabled by default at -Os, -O2, and -O3.

Constant propagation

Functions where all callers pass the same constant can be optimized by propagating the constant into the function. That is,

static int foo(int a, int b)
{
  if (b > 0)
    return a + b;
  else
    return a * b;
}

int bar(int m, int n)
{
  return foo(m, 7) + foo(n, 7);  
}

is optimized to

static int foo(int a)
{
  return a + 7;
}

int bar(int m, int n)
{
  return foo(m) + foo(n);
}

The constants can be propagated bitwise, which is useful for flag parameters. For example

static int foo(int a, int b)
{
  if (b & 4)
    return a & (b & 1);
  else
    return a & (b & 2);
}

int bar(int m, int n)
{
  return foo(m, 9) | foo(n, 3);
}

is optimized to

static int foo(int a, int b)
{
  return a & (b & 2);
}

int bar(int m, int n)
{
  return foo(m, 9) | foo(n, 3);
}

The constants do not need to be the same in all function calls – GCC tracks ranges of possible values and optimize as appropriate, so

static int foo(int a, int b)
{
  if (b > 0)
    return a + b;
  else
    return a * b;
}

int bar(int m, int n)
{
  return foo(m, 5) + foo(n, 7);  
}

is optimized to

static int foo(int a, int b)
{
  return a + b;
}

int bar(int m, int n)
{
  return foo(m, 5) + foo(n, 7);  
}

as both 5 and 7 are greater than 0.

These optimizations are enabled by -fipa-cp, -fipa-bit-cp, and -fipa-vrp, which are enabled by default at -Os, -O2, and -O3.

Constant propagation – cloning

It is often the case that only a few of the function calls pass constants as parameters, or that the constants are conflicting so they cannot be propagated into the called function. GCC handles this by cloning the called function to let each conflicting call get its own version. For example,

static int foo(int a, int b)
{
  if (b > 0)
    return a + b;
  else
    return a * b;
}

int bar(int m, int n)
{
  return foo(m, 5) + foo(m, n);  
}

creates one clone of foo and optimizes it using the constant 5 for the parameter b

static int foo(int a, int b)
{
  if (b > 0)
    return a + b;
  else
    return a * b;
}

static int foo_clone(int a)
{
  return a + 5;
}

int bar(int m, int n)
{
  return foo_clone(m) + foo(m, n);  
}

This optimization is enabled by -fipa-cp-clone, which is enabled by default at -O3.

Devirtualization

Devirtualization (converting calls to virtual functions to direct calls – see Jan Hubička's blog series on how devirtualization works in GCC) is helped by propagating type information in roughly the same way as the constants are propagated, and is implemented by the constant propagation pass.

This is enabled by -fipa-cp and -fdevirtualize, which are enabled by default at -Os, -O2, and -O3.

Caller-saved registers

Caller saved registers do not need to be saved if those registers are not used by the called function.

This optimization is enabled by -fipa-ra, which is enabled by default at -Os, -O2, and -O3.

Identical code folding

The “identical code folding pass” merges identical functions. The functions do not need to be identical in the source code – the merging is done halfway through the optimization pipeline so it is enough that they have the same structure after simplification (and variable names etc. does not matter).

Functions that may be used outside the compilation unit cannot be completely merged as the C and C++ standards require that functions have unique addresses. GCC solves this by adding wrappers for the exported symbols, so that

#include <stdio.h>

void foo(char *s)
{
    printf("Hello %s\n", s);
}

void bar(char *s)
{
    printf("Hello %s\n", s);
}

is generated as

.LC0:
        .string "Hello %s\n"

foo:
        mov     rsi, rdi
        xor     eax, eax
        mov     edi, OFFSET FLAT:.LC0
        jmp     printf

bar:
        jmp     foo

This optimization is enabled by -fipa-icf, which is enabled by default at -Os, -O2, and -O3.

Profile propagation

Many optimizations have different heuristics depending on how much the code is executed. The compiler estimates branch frequencies and propagates this information between functions so that, for example, a function only called from “cold” code segments is treated as a “cold” function.

This is enabled by -fipa-profile, which is enabled by default at -O and higher.

Pure, const, and noexcept

GCC analyzes functions to determine if they access memory or may throw exceptions, propagates this information throughout the compilation unit, and annotates the functions with pure, const, and noexcept attributes when possible, which helps other optimizations.

This optimization is enabled by -fipa-pure-const, which is enabled by default at -O and higher.

Global variables

It is in general hard to optimize usage of global variables, but it is easy to improve usage of global variables that cannot escape the compilation unit and that do not have the address taken. There are three optimizations done on such variables

• Removal of global variables that are never read.
• A global variable that is used in only one function may be changed to a local variable in that function.
• The compiler tracks which functions modifies the variables so that loads and stores may be moved over function calls that do not touch the variable. For example, the function bar in

  static int g;
  
  void foo(void)
  {
    // Code not touching g
  }
  
  int bar(void)
  {
    g += 1;
    foo();
    g += 2;
  }
  

  is optimized to

  int bar(void)
  {
    foo();
    g += 3;
  }
  

These optimizations are enabled by -fipa-reference, which is enabled by default at -O and higher.

Pointer analysis

GCC can do interprocedural pointer analysis, which is enabled by -fipa-pta. This optimization is not enabled by default at any optimization level as it can cause excessive memory and compile-time usage on large compilation units.

Seeding the std::mt19937 random number engine

A comment on Hacker News complained that the code in my previous blog post does not seed the std::mt19937 random number engine properly. The code was taken directly from a CppCon presentation, so I don’t want to take the blame, but the comment is right — the initialization code can be improved.

State size and seeding

The initialization in the blog post was done as

std::random_device rd;
std::mt19937 gen(rd());

which seeds the std::mt19937 random number engine with a random 32-bit value. The problem with this is that that the Mersenne twister has 19968 bits of internal state so it can generate \(2^{19968}\) streams of random values, but we can only reach \(2^{32}\) of those states when initializing with a 32-bit value.

This is not necessarily a problem. Let’s say the random numbers are used for generating input data in unit tests. The test suite is probably not run more than a few thousand times, so it does not matter that it only can create \(2^{32}\) different test runs. But there are use-cases where this is a problem.

The random number engine can be seeded with more data by using std::seed_seq, and the code below seeds the std::mt19937 with the same number of bits as are in the state

std::random_device rd;
std::array<int, std::mt19937::state_size> seed_data;
std::generate_n(seed_data.data(), seed_data.size(), std::ref(rd));
std::seed_seq seq(std::begin(seed_data), std::end(seed_data));
std::mt19937 gen(seq);

std::random_device

One other potential problem is the quality of the seed values. The idea behind std::random_device is that it returns non-deterministic random numbers, but it is allowed to return deterministic values (e.g. if a non-deterministic source is not available to the implementation). I’m not a big fan of this functionality — it either does exactly what you want (generates non-deterministic values) or it does the opposite (generates deterministic values), and there is no way you can determine which.¹

This is probably not a problem when developing for the big platforms, but there may be surprises when running the code in other environments — at least old versions of libstdc++ on MinGW always return the same sequence of values...

---

1. The std::random_device can return an estimate of the entropy, and it is required to return 0 if the values are generated deterministically. But it is not required to return non-zero for the non-deterministic case, and e.g. libstdc++ is conservative and always estimates the entropy as 0, even when /dev/urandom or the x86 RDRND instruction are used.

Integer division is slow

The CppCon 2016 talk “I Just Wanted a Random Integer!” benchmarks randomization functionality from the C++ standard library (using GCC 5.1). There is one surprising result — the loop

std::random_device entropySource;
std::mt19937 randGenerator(entropySource());
std::uniform_int_distribution<int> theIntDist(0, 99);

for (int i = 0; i < 1000000000; i++) {
  volatile auto r = theIntDist(randGenerator);
}

need 23.4 seconds to run while

std::random_device entropySource;
std::mt19937 randGenerator(entropySource());

for (int i = 0; i < 1000000000; i++) {
  std::uniform_int_distribution<int> theIntDist(0, 99);
  volatile auto r = theIntDist(randGenerator);
}

run in 5.1 seconds. But the latter should intuitively be slower as it does more in the loop...

Code expansion

The functionality in the standard library is implemented using template magic, but the compiler’s view of the code after inlining and basic simplification is that

std::uniform_int_distribution<int> theIntDist(0, 99);

is just defining and initializing a structure

struct {
  int a, b;
} theIntDist;
theIntDist.a = 0;
theIntDist.b = 99;

while the call

volatile auto r = theIntDist(randGenerator);

is expanded to the equivalent of

uint64_t ret;
uint64_t urange = theIntDist.b - theIntDist.a;
if (0xffffffff > urange) {
  const uint64_t uerange = urange + 1;
  const uint64_t scaling = 0xffffffff / uerange;
  const uint64_t past = uerange * scaling;
  do {
    ret = mersenne_twister_engine(randGenerator);
  } while (ret >= past);
  ret /= scaling;
} else {
  ...
  uniform_int_distribution(&theIntDist, randGenerator);
  ...
  ret = ...
}
volatile int r = ret + theIntDist.a;

where I have used ... for code that is not relevant for the rest of the discussion.

Optimization differences

It is now easy to see why the second case is faster — creating theIntDist in the loop makes it trivial for the compiler to determine that urange has the value 99, and the code simplifies to

uint64_t ret;
do {
  ret = mersenne_twister_engine(randGenerator);
} while (ret >= 4294967200);
ret /= 42949672;
volatile int r = ret;

This simplification is not possible when theIntDist is created outside of the loop — the compiler sees that the loop calls uniform_int_distribution with a reference to theIntDist, so it must assume that the value of theIntDist.a and theIntDist.b may change during the execution and can therefore not do the constant folding. The function does, however, not modify theIntDist, so both versions of the program do the same work, but the slow version needs to do one extra comparison/branch and a few extra arithmetic instructions for each loop iteration.

The cost of division

The mersenne_twister_engine is not a big function, but it is not trivial — it executes about 40 instructions — so it is surprising that adding a few instructions to the loop makes the program four times slower. I described a similar case in a previous blog post where the problem were due to branch mis-prediction, but the branch is perfectly predicted in  this example.

The reason here is that the slow loop need to do an integer division instruction when calculating scaling, and integer division is expensive — Agner Fog’s instruction tables says that the 64-bit division may need up to 103 cycles on the Broadwell microarchitecture! This usually does not matter too much for normal programs as as the compiler tries to move the division instructions so that they have as much time as possible to execute before the result is needed, and the CPU can in general continue executing other instructions out of order while waiting for the result of the division. But it does make a big difference in this kind of micro-benchmarks as the compiler cannot move the division earlier, and the CPU runs out of work to do out of order as the mersenne_twister_engine function executes much faster than the division.

Building GCC with support for NVIDIA PTX offloading

GCC can offload C, C++, and Fortran code to an accelerator when using OpenACC or OpenMP where the code to offload is controlled by adding #pragma statements (or magic comments for Fortran), such as

#pragma acc kernels
for (int j = 1; j < n-1; j++) {
  for (int i = 1; i < m-1; i++) {
    Anew[j][i] = 0.25f * (A[j][i+1] + A[j][i-1] + A[j-1][i] + A[j+1][i]);
    error = fmaxf(error, fabsf(Anew[j][i] - A[j][i]));
  }
}

This blog post describes what I needed to do in order to build a GCC 7.2 compiler with support for offloading to NVIDIA GPUs on Ubuntu 16.10.

The first step is to install the NVIDIA CUDA toolkit. Googling shows lots of strange suggestions about what you need to do in order to get this to work (blacklisting drivers, adding the PCI address of your video card to config files, etc.), but it worked fine for me to just download the “deb (local)” file, and install it as

sudo dpkg -i cuda-repo-ubuntu1604-8-0-local-ga2_8.0.61-1_amd64.deb
sudo apt-get update
sudo apt-get install cuda

The toolkit is installed in /usr/local/cuda, and /usr/local/cuda/bin must be added to PATH so that GCC may find the ptxas tool.

The script below fetches the source code and builds the compiler and tools

Add $install_dir/lib64 to LD_LIBRARY_PATH, and the compiler can now be used to offload OpenACC code by compiling as

$install_dir/bin/gcc -O3 -fopenacc test.c

or OpenMP as

$install_dir/bin/gcc -O3 -fopenmp test.c

You may need to pass -foffload=-lm to the compiler if the code you offload contains math functions that cannot be directly generated as PTX instructions.

Updated 2017-12-23: Changed the script to build GCC 7.2 instead of trunk as there are some problems with the trunk compiler at the moment... 

Updated 2021-05-01: Update the script to build GCC 11 and a newer version of newlib in order to solve build issues on newer versions of Ubuntu.

pre-decrement vs. post-decrement, etc.

A recent talk at the OpenIoT Summit NA 2017, “Optimizing C for Microcontrollers — Best Practices”, had three examples illustrating the effect of different code constructs

• Array subscript vs. pointer access
• Loops (increment vs. decrement)
• Loops (post-decrement vs. pre-decrement)

as compiled using GCC 6.x on ARM and the -Os optimization level. This blog post will look a bit closer at those examples, and discuss why the conclusions are not always valid.

Array subscript vs. pointer access

The first example is meant to illustrate the difference between array subscripts and pointer access with the two functions

int a[5];

int foo1(void)
{
    int i;
    int res = 0;
    for (i = 0; i < 5; i++)
        res += a[i];
    return res;
}

and

int a[5];

int foo2(void)
{
    int *p;
    int i;
    int res = 0;
    for (p = a, i = 0; i < 5; i++, p++)
        res += *p;
    return res;
}

The first function is generated in a natural way

foo1:
    movs  r0, #0
    mov   r3, r0
    ldr   r1, .L5
.L3:
    ldr   r2, [r1, r3, lsl #2]
    adds  r3, r3, #1
    cmp   r3, #5
    add   r0, r0, r2
    bne   .L3
    bx    lr

while the second function has its loop unrolled

foo2:
    ldr   r3, .L3
    ldm   r3, {r0, r2}
    add   r0, r0, r2
    ldr   r2, [r3, #8]
    add   r0, r0, r2
    ldr   r2, [r3, #12]
    ldr   r3, [r3, #16]
    add   r0, r0, r2
    add   r0, r0, r3
    bx    lr

The reason for this difference is that compiling with -Os should not unroll loops if unrolling increases the code size. But it is hard to estimate the resulting code size, as later optimization passes should be able to take advantage of the unrolling and be able to remove redundant code, so the compiler is using a rather imprecise heuristic. These loops are really close to the threshold (unrolling increases the code size by 4 bytes) and the minor difference between how the loops look when passed to the unroller makes the heuristic estimate that unrolling foo1 will increase the size by one instruction while foo2 will get the same size after unrolling.

This does, however, not illustrate any fundamental difference in the compiler’s understanding of array subscript compared pointer access — any difference in the code could affect a heuristic and have a similar effect (I have worked on compilers that generate different code if you rename variables or even add a comment!).¹

Loops (increment vs. decrement)

The second example uses the two functions

void foo1(void)
{
    int x = 0;
    do {
        printk("X = %d\n", x);
        x++;
    } while (x < 100);
}

and

void foo2(void)
{
    int x = 100;
    do {
        printk("X = %d\n", x);
        x--;
    } while (x);
}

to illustrate that it is better to write loops decrementing the iteration variable, as the CPU can do the end of loop check for free as

subs  r4, r4, #1
bne   .L3 

instead of

adds  r4, r4, #1
cmp   r4, #100
bne   .L3

That is true, but the compiler can in many cases transform the loop to change iteration order, so the iteration order in the generated program depend more on what the loop does than how it iterates in the source code.

Note that the two functions do not do the same thing — foo1 outputs the numbers in increasing order and foo2 outputs them in decreasing order. Modifying foo2 to do the same thing as foo1, by changing the function call to

printk("X = %d\n", 100 - x);

makes it generate identical code as foo1 (as the compiler decides that it is better to iterate using increments in order to eliminate the subtraction) even though the function was written as using decrements.

Loops (post-decrement vs. pre-decrement)

The third example consider pre- vs. post-decrement using the examples

void foo1(void)
{
    unsigned int x = 10;
    do {
        if (--x) {
            printk("X = %d\n", x);
        } else {
            printk("X = %d\n", x);
            x = 10;
        }
    } while (1);
}

and

void foo2(void)
{
    unsigned int x = 9;
    do {
        if (x--) {
            printk("X = %d\n", x);
        } else {
            printk("X = %d\n", x);
            x = 9;
        }
    } while (1);
}

The example is meant to illustrate that --x is better, as it can get the comparison as a side effect of the subtraction in the same way as the previous example

subs  r4, r4, #1
bne   .L3 

but it depends much on the microarchitecture if this is beneficial or not. Many microarchitectures can do compare and branch efficiently,² so a compare and a branch are not necessarily slower than branching on the status code from the subtraction. The problem with --x is that it adds a data dependency — you must do the subtraction before you can evaluate the if-statement. With x-- you can evaluate the if-statement and subtraction in parallel, with the result that

if (--x)

need one extra cycle to execute compared to

if (x--)

for superscalar CPUs having efficient compare and branch.

---

1. This typically happens when the compiler has different equivalent choices (for example, should it spill variable a or b to the stack), and it just chooses the first alternative. The first alternative is found by iterating over some kind of container, and this container may be an associative array using pointer values as the key...
2. For example, x86 CPUs tend to fuse cmp and jne so that they execute as one instruction.