C Structure Padding Initialization

This article takes a look at a few different aspects of C structure initialization. In particular, we’ll look at when it matters, the current state of things in Clang and GCC, recommendations, and the ✨ future ✨.

Time to dive into this very niche, but occasionally hazardous corner of the C language!

C Structure Padding

Recently I was reading this excellent post on some of the upcoming features in C23, and it inspired me to do a little exploration and documentation around the current state of initialization of padding in structures in the C-language.

For background, let’s take this example C structure:

#include <stdint.h>

struct foo {
  uint32_t i;
  uint8_t b;
};

By default, padding will be inserted at the end of this structure, to align it to the largest member size. We can use the pahole tool to examine structure holes after compiling (with debug symbols enabled, -g):

struct foo {
        uint32_t                   i;                    /*     0     4 */
        uint8_t                    b;                    /*     4     1 */

        /* size: 8, cachelines: 1, members: 2 */
        /* padding: 3 */
        /* last cacheline: 8 bytes */
};

My understanding is this is done so if the structure is addressed as part of an array, the first member of each element in the array will have the same alignment:

struct foo foo_array[2];
// if &foo_array[0] has 4 byte alignment, we want &foo_array[1]
// to also have 4 byte alignment. This is because on most architectures it is more efficient to access data along boundaries aligned with their size.

The other (more commonly encountered) case is where padding is inserted between structure members:

struct foo_internal_padding {
  uint8_t b;
  uint32_t i;
};

Running pahole as above shows the padding inserted after the first element:

struct foo {
        uint8_t                    b;                    /*     0     1 */

        /* XXX 3 bytes hole, try to pack */

        uint32_t                   i;                    /*     4     4 */

        /* size: 8, cachelines: 1, members: 2 */
        /* sum members: 5, holes: 1, sum holes: 3 */
        /* last cacheline: 8 bytes */
};

Note! the excellent guide here (which we’ve linked before), is a great reference on structure padding and artisanal hand-packing: http://www.catb.org/esr/structure-packing/

Of course, to prevent padding, we can force the compiler to pack the structure:

struct __attribute__((packed)) foo {
  uint8_t b;
  uint32_t i;
};

Now we have no padding inside the structure:

struct foo {
        uint8_t                    b;                    /*     0     1 */
        uint32_t                   i;                    /*     1     4 */

        /* size: 5, cachelines: 1, members: 2 */
        /* last cacheline: 5 bytes */
} __attribute__((__packed__));

Similarly, a structure that would normally have padding at the end will no longer have it:

struct foo {
        uint32_t                   i;                    /*     0     4 */
        uint8_t                    b;                    /*     4     1 */

        /* size: 5, cachelines: 1, members: 2 */
        /* last cacheline: 5 bytes */
} __attribute__((__packed__));

Note that accessing members of compiler-packed structs often can add compute overhead; the CPU may need to do bytewise loads and stores depending on alignment requirements of the architecture.

For completeness, note that arrays of packed structures by default will also be packed (no trailing padding inserted between array elements; example here). Usually the compiler will do the correct thing, but you might run into unexpected cases when type aliasing (this is undefined behavior anyway, and there be dragons here 🐉!).

Bitfields follow similar rules when it comes to packing, with the added complexity where the type holding the bitfield is undefined, with this somewhat horrifying language in the C11 specification §6.7.2.1/11:

An implementation may allocate any addressable storage unit large enough to hold a bit-field. If enough space remains, a bit-field that immediately follows another bit-field in a structure shall be packed into adjacent bits of the same unit. If insufficient space remains, whether a bit-field that does not fit is put into the next unit or overlaps adjacent units is implementation-defined. The order of allocation of bit-fields within a unit (high-order to low-order or low-order to high-order) is implementation-defined. The alignment of the addressable storage unit is unspecified.

As Mitch Johnson over at theinterrupt.slack.com pointed out, there are other subtleties to consider with bitfields that can have architecture-specific implications:

… some architectures (ARM in particular) require compilers to represent volatile bitfield layout and accesses in a well-defined fashion in order to comply with their procedure call standard. This allows use of volatile bitfields to properly represent access to memory-mapped peripherals. https://developer.arm.com/documentation/ihi0042/j/?lang=en This can still be fraught with danger. GCC’s had a number of bugs around volatile bitfield usage, and ARM’s own clang derivative has had varyingly non-compliant behavior over time: https://developer.arm.com/documentation/ka004594/latest

When does this matter?

Alright! Now that we’ve got a description of struct padding, let’s describe some cases where it makes a difference.

Comparing padded structs

It’s tempting to compare structs by using memcmp, as in the following:

struct foo {
  uint32_t i;
  uint8_t b;
};

// Check 2 foos for equality
bool foo_are_equal(struct foo a, struct foo b) {
  const int result = memcmp(&a, &b, sizeof(a));
  return result == 0;
}

// Check if a foo matches a reference
bool foo_is_reference(struct foo a) {
  static const struct foo reference = {
    .i = 1234,
    .b = 56,
  };
  return foo_are_equal(a, reference);
}

HOWEVER, this may give incorrect results if the padding is not accounted for!

Serializing structs outside the application

For example, writing a C struct into non-volatile storage:

struct device_config {
  uint8_t device_config_version;
  uint64_t manufacture_date;
  uint8_t hardware_version;
  uint8_t serial_number[16];
};

If that data needs to be read by another piece of software, or if it potentially will be migrated to a different struct layout (eg, a new field is added), it might be prudent to pack the struct (either by hand or with __attribute__((packed))), to simplify reasoning about the data structure.

Security issues

The values in the padding space can potentially leak sensitive information if the data structures are crossing trust boundaries. Specifically, the padding space can contain data from objects that were previously allocated on the stack (for example, an encryption key used to perform some cryptographic operation).

See the following articles for information on that subject:

When it doesn’t matter

If the structure is only ever accessed on a per-member basis, the padding probably won’t cause problems:

struct foo {
  uint32_t i;
  uint8_t b;
};

// Check 2 foos for equality
bool foo_are_equal(struct foo a, struct foo b) {
  return (a.i == b.i) && (a.b == b.b);
}

Specifically, if the structure is only ever internally used in the application (never crosses a library or trust boundary, and is never serialized out to external storage or over a communications interface), issues related to padding may not be a problem.

Note that other languages may have different layout implementations for composite types (see Rust), which may complicate matters when moving raw C structs between different pieces of software.

Structure (Zero) Initialization

Given the above, it seems convenient to zero-initialize structures before using them. With C99 or later, it is common to make use of the following patterns with “designated initializers” for structure initialization:

struct foo {
  uint32_t i;
  uint8_t b;
};

// Initialize members of 'a' to specific values. Members not specifically
// initialized will be initialized per the 'static storage duration'
// initialization rules (eg pointers go to NULL, integers go to 0, floats go to
// 0.0, etc)
struct foo a = {
  .i = 1,
  // .b will be set to 0
};

// Initialize 'b' to all zeros. This is a common idiom that specifies a '0'
// constant as the initial value for the first member of the structure, then
// relies on the above rule to initialize the rest of the structure.
struct foo b = { 0 };

This looks great! However, it’s not obvious (from looking at those snippets) what the value loaded into the padding region will be.

The unfortunate answer is: it depends

The C11 standard, chapter §6.2.6.1/6 says this:

When a value is stored in an object of structure or union type, including in a member object, the bytes of the object representation that correspond to any padding bytes take unspecified values.

See also https://stackoverflow.com/a/37642061

Objects with ‘static storage duration’ (static keyword, or external linkage (defined at the outermost scope in a compilation unit)), padding bits will be initialized to 0!

Objects with ‘automatic storage duration’ (locally-scoped objects) have undefined behavior when it comes to padding bit initialization!

This means that there is no constraint on what values are set to those bits when the object is initialized.

The Current State

Let’s consider the following 4 zero-initialization strategies for this structure:

struct foo {
  uint32_t i;
  uint8_t b;
};
  1. memset to zeros:

    struct foo a;
    memset(&a, 0, sizeof(a));
    
  2. individually set all members to 0:

    struct foo a = {
      .i = 0,
      .b = 0,
    };
    
  3. use { 0 } zero-initializer

    struct foo a = { 0 };
    
  4. use {} GCC extension zero-initializer (Note: this is quite poorly/non-documented for C - it IS valid C++ - but works in C on both GCC and clang. See here and here)

    struct foo a = {};
    

It turns out, the results for these vary between GCC and Clang and optimization levels.

For the record, I’m testing using these compiler versions, on Ubuntu Linux 21.10 on 2022-02-28:

❯ clang --version
Ubuntu clang version 13.0.0-2

❯ gcc --version
gcc (Ubuntu 11.2.0-7ubuntu2) 11.2.0

# specific package versions:
❯ apt list clang gcc
Listing... Done
clang/impish,now 1:13.0-53~exp1 amd64 [installed]
gcc/impish,now 4:11.2.0-1ubuntu1 amd64 [installed]

Padding values under each strategy, optimization level, and compiler (warning boring tables below!):

Strategy 1, memset

Strategy Optimization Level Clang 13 GCC 11
1, memset 0 zero zero
1, memset 1 zero zero
1, memset 2 zero zero
1, memset 3 zero zero
1, memset s zero zero

Strategy 2, explicitly setting each struct member

Strategy Optimization Level Clang 13 GCC 11
2, explicit 0 zero unset
2, explicit 1 unset unset
2, explicit 2 zero unset
2, explicit 3 zero unset
2, explicit s unset unset

Strategy 3, { 0 }

Strategy Optimization Level Clang 13 GCC 11
3, { 0 } 0 zero zero
3, { 0 } 1 unset zero
3, { 0 } 2 zero zero
3, { 0 } 3 zero zero
3, { 0 } s zero zero

Strategy 4, { } GCC extension

Strategy Optimization Level Clang 13 GCC 11
4, { } 0 zero zero
4, { } 1 unset zero
4, { } 2 zero zero
4, { } 3 zero zero
4, { } s unset zero

The main point is that it’s not particularly consistent across compilers and optimization levels 😱!

You can find the example application used to generate the above data here on Github.

I’ve also uploaded it to the wonderful Compiler Explorer if you want to take a look and quickly play around: https://godbolt.org/z/b985G4ejT

Best Practice?

It’s tricky to recommend a one-size-fits-all option here, because different software will have different constraints. However, some general purpose advice follows.

Avoid relying on structure layout

The simplest option to avoid padding issues is to avoid the padding fields altogether:

Access data directly via each member, do not alias structures or use memcmp etc.

This only works if the padding data can be safely ignored in all use cases for the data structures in question.

Note however that packed structs can be safely memcmp‘d, see below

Another approach is to avoid structure holes entirely!

For example, you can use the -Wpadded compiler warning in GCC and Clang to detect padding, and with -Werror or -Werror=padded, you can trigger compilation errors if padding is detected. To address the warnings, you can add placeholders to fill unused space:

struct foo {
  uint32_t i;
  uint8_t b;
  uint8_t padding_[3];
};

(Note that GCC will emit a warning on declaration, where Clang will only warn when the violating struct is actually used in a definition. Similar, but subtly different as usual 🌈).

Alternatively, the pahole tool could be used to detect any structures with padding bits (for example, as a linter pass on the generated binary), and they can be corrected by either reordering structure members to eliminate padding, or adding uint8_t padding_[n] fields to explicitly address the holes. See also The Lost Art of Structure Packing linked previously.

It’s generally preferable to strive for padding to only be present at end of struct.

Use memset to zero-initialize padding bits

memset reliably sets the entire memory space of an object, including the padding bits.

Must be manually done, though, so can be error-prone.

Be sure to set the size argument directly from the object in question:

// error prone! if the type of 'a' changes, we might get unexpected results
memset(&a, 0, sizeof(struct foo));

// much better
memset(&a, 0, sizeof(a));

Last resort, __attribute__((packed))

This will eliminate structure padding, but there can be considerable compute overhead (and with code doing unusual type aliasing, you may find yourself in an Unaligned Access fault 😓).

On the plus side, the structure fields should have easily predictable offsets in memory, for example if it needs to be serialized out.

Additionally, packed structs can be safely memcmp‘d, since there are no “ghost” bits hiding in between explicitly allocated members 👻

The Future 🌞

I gave this away right at the start of the article, but somewhat unusually for C, there is change coming on this topic!

The proposed change for C23 is that the = {} (functionally equivalent to = { 0 } on modern compilers) will also initialize padding bits to 0 😎.

You can see the gory details in the following links:

This seems like a nice update to the standard that doesn’t appear to impact backwards-compatibility and just makes things better!

Update 2022-05-19

GCC 12 has added a new flag, -ftrivial-auto-var-init=choice which enables zero-initialization of struct padding by the compiler:

https://gcc.gnu.org/onlinedocs/gcc-12.2.0/gcc/Optimize-Options.html#index-ftrivial-auto-var-init

This flag has been present in Clang for a while, but it seems like it may be removed in the future; it requires setting this somewhat amusingly named flag to use:

-enable-trivial-auto-var-init-zero-knowing-it-will-be-removed-from-clang

The flag is documented here.

This flag definitely incurs runtime penalties, so it might make sense to enable it when compiling some sensitive subset of your application (eg a security library). Definitely an interesting compiler feature!

See anything you'd like to change? Submit a pull request or open an issue on our GitHub

References

Noah Pendleton is an embedded software engineer at Memfault. Noah previously worked on embedded software teams at Fitbit and Markforged