Minimizing Binary Size

[中文]

The ESP-IDF build system compiles all source files in the project and ESP-IDF, but only functions and variables that are actually referenced by the program are linked into the final binary. In some cases, it is necessary to reduce the total size of the firmware binary, e.g., in order to fit it into the available flash partition size.

The first step to reducing the total firmware binary size is measuring what is causing the size to increase.

Measuring Static Sizes

To optimize both the firmware binary size and the memory usage, it is necessary to measure statically-allocated RAM (data, bss), code (text), and read-only data (rodata) in your project.

Using the idf.py sub-commands size, size-components, and size-files provides a summary of memory used by the project:

Size Summary idf.py size

$ idf.py size
[...]
Total sizes:
Used stat D/IRAM:   53743 bytes ( 122385 remain, 30.5% used)
      .data size:    6504 bytes
      .bss  size:    1984 bytes
      .text size:   44228 bytes
   .vectors size:    1027 bytes
Used Flash size :  118879 bytes
           .text:   83467 bytes
         .rodata:   35156 bytes
Total image size:  170638 bytes (.bin may be padded larger)

This output breaks down the size of all static memory regions in the firmware binary:

$ idf.py size
[...]
Total sizes:
Used stat D/IRAM:   53743 bytes ( 122385 remain, 30.5% used)
      .data size:    6504 bytes
      .bss  size:    1984 bytes
      .text size:   44228 bytes
   .vectors size:    1027 bytes
Used Flash size :  118879 bytes
           .text:   83467 bytes
         .rodata:   35156 bytes
Total image size:  170638 bytes (.bin may be padded larger)
  • Used stat D/IRAM: Total amount of D/IRAM used at compile time. remain indicates the amount of D/IRAM left to be used as heap memory at runtime. Note that due to meta data overhead, implementation constraints, and startup heap allocations, the actual size of the DRAM heap is smaller.

    • .data size: Amount of D/IRAM allocated at compile time for the .data (i.e., all statically allocated variables that are initialized to non-zero values). .data also consumes space in the binary image to store the non-zero initialization values.

    • .bss  size: Amount of D/IRAM allocated at compile time for .bss (i.e., all statically allocated variables that are initialized to zero). .bss does not consume extra space in flash.

    • .text size: Amount of D/IRAM used for .text (i.e., all code that is executed from internal RAM). .text also consumes space in the binary image as the code is initially stored there and is then copied over to D/IRAM on startup.

  • Used Flash size: Total amount of flash used (excluding usage by D/IRAM)

    • .text: Amount of flash used for .text (i.e., all code that is executed via the flash cache, see IROM).

    • .rodata: Amount of flash used for .rodata (i.e., read-only data that is loaded via the flash cache, see DROM).

  • Total image size is the estimated total size of the binary file.

Component Usage Summary idf.py size-components

The summary output provided by idf.py size does not give enough details to find the main contributor to excessive binary size. To analyze in detail, use idf.py size-components.

$ idf.py size-components
[...]
    Total sizes:
 DRAM .data size:   14956 bytes
 DRAM .bss  size:   15808 bytes
Used static DRAM:   30764 bytes ( 149972 available, 17.0% used)
Used static IRAM:   83918 bytes (  47154 available, 64.0% used)
      Flash code:  559943 bytes
    Flash rodata:  176736 bytes
Total image size:~ 835553 bytes (.bin may be padded larger)
Per-archive contributions to ELF file:
            Archive File DRAM .data & .bss & other   IRAM   D/IRAM Flash code & rodata   Total
           libnet80211.a       1267   6044       0   5490        0     107445    18484  138730
               liblwip.a         21   3838       0      0        0      97465    16116  117440
            libmbedtls.a         60    524       0      0        0      27655    69907   98146
         libmbedcrypto.a         64     81       0     30        0      76645    11661   88481
                 libpp.a       2427   1292       0  20851        0      37208     4708   66486
                  libc.a          4      0       0      0        0      57056     6455   63515
                libphy.a       1439    715       0   7798        0      33074        0   43026
     libwpa_supplicant.a         12    848       0      0        0      35505     1446   37811
           libfreertos.a       3104    740       0  15711        0        367     4228   24150
          libnvs_flash.a          0     24       0      0        0      14347     2924   17295
          libspi_flash.a       1562    294       0   8851        0       1840     1913   14460
         libesp_system.a        245    206       0   3078        0       5990     3817   13336
            libesp-tls.a          0      4       0      0        0       5637     3524    9165
[... removed some lines here ...]
            libesp_rom.a          0      0       0    112        0          0        0     112
                libcxx.a          0      0       0      0        0         47        0      47
                   (exe)          0      0       0      3        0          3       12      18
             libesp_pm.a          0      0       0      0        0          8        0       8
            libesp_eth.a          0      0       0      0        0          0        0       0
               libmesh.a          0      0       0      0        0          0        0       0

The first lines of the output from idf.py size-components are the same as that from idf.py size. After this, a table is printed as Per-archive contributions to ELF file. This means how much each static library archive has contributed to the final binary size.

Generally, one static library archive is built per component, although some are binary libraries included by a particular component, for example, libnet80211.a is included by esp_wifi component. There are also toolchain libraries such as libc.a and libgcc.a listed here, these provide Standard C/C++ Library and toolchain built-in functionality.

If your project is simple and only has a main component, then all of the project's code will be shown under libmain.a. If your project includes its own components (see Build System), then they will each be shown on a separate line.

The table is sorted in descending order of the total contribution of the static archive to the binary size.

The columns are as follows:

  • DRAM .data & .bss & other - .data and .bss are the same as for the totals shown above. Both are static variables and reduce the total available RAM at runtime, but .bss does not contribute to the binary file size. other is a column for any custom section types that also contribute to RAM size. Usually, the value is 0.

  • IRAM - is the same as for the totals shown above. It refers to code linked to execute from IRAM, which uses space in the binary file and also reduces DRAM available as heap at runtime.

  • Flash code & rodata - these are the same as the totals above, IROM and DROM space accessed from the flash cache that contribute to the binary size.

Source File Usage Summary idf.py size-files

For even more details, run idf.py size-files to get a summary of the contribution each object file has made to the final binary size. Each object file corresponds to a single source file.

$ idf.py size-files
[...]
Total sizes:
 DRAM .data size:   14956 bytes
 DRAM .bss  size:   15808 bytes
Used static DRAM:   30764 bytes ( 149972 available, 17.0% used)
Used static IRAM:   83918 bytes (  47154 available, 64.0% used)
      Flash code:  559943 bytes
    Flash rodata:  176736 bytes
Total image size:~ 835553 bytes (.bin may be padded larger)
Per-file contributions to ELF file:
             Object File DRAM .data & .bss & other   IRAM   D/IRAM Flash code & rodata   Total
     x509_crt_bundle.S.o          0      0       0      0        0          0    64212   64212
                wl_cnx.o          2   3183       0    221        0      13119     3286   19811
           phy_chip_v7.o        721    614       0   1642        0      16820        0   19797
       ieee80211_ioctl.o        740     96       0    437        0      15325     2627   19225
                    pp.o       1142     45       0   8871        0       5030      537   15625
      ieee80211_output.o          2     20       0   2118        0      11617      914   14671
         ieee80211_sta.o          1     41       0   1498        0      10858     2218   14616
        lib_a-vfprintf.o          0      0       0      0        0      13829      752   14581
       lib_a-svfprintf.o          0      0       0      0        0      13251      752   14003
             ssl_tls.c.o         60      0       0      0        0      12769      463   13292
             sockets.c.o          0    648       0      0        0      11096     1030   12774
                 nd6.c.o          8    932       0      0        0      11515      314   12769
       phy_chip_v7_cal.o        477     53       0   3499        0       8561        0   12590
                    pm.o         32    364       0   2673        0       7788      782   11639
        ieee80211_scan.o         18    288       0      0        0       8889     1921   11116
      lib_a-svfiprintf.o          0      0       0      0        0       9654     1206   10860
       lib_a-vfiprintf.o          0      0       0      0        0      10069      734   10803
          ieee80211_ht.o          0      4       0   1186        0       8628      898   10716
       phy_chip_v7_ana.o        241     48       0   2657        0       7677        0   10623
              bignum.c.o          0      4       0      0        0       9652      752   10408
              tcp_in.c.o          0     52       0      0        0       8750     1282   10084
                   trc.o        664     88       0   1726        0       6245     1108    9831
               tasks.c.o          8    704       0   7594        0          0     1475    9781
          ecp_curves.c.o         28      0       0      0        0       7384     2325    9737
                 ecp.c.o          0     64       0      0        0       8864      286    9214
      ieee80211_hostap.o          1     41       0      0        0       8578      585    9205
                  wdev.o        121    125       0   4499        0       3684      580    9009
             tcp_out.c.o          0      0       0      0        0       5686     2161    7847
                 tcp.c.o          2     26       0      0        0       6161     1617    7806
       ieee80211_input.o          0      0       0      0        0       6797      973    7770
                 wpa.c.o          0    656       0      0        0       6828       55    7539
[... additional lines removed ...]

After the summary of total sizes, a table of Per-file contributions to ELF file is printed.

The columns are the same as shown above for idy.py size-components, but this time the granularity is the contribution of each individual object file to the binary size.

For example, we can see that the file x509_crt_bundle.S.o contributed 64,212 bytes to the total firmware size, all as .rodata in flash. Therefore we can guess that this application is using the ESP x509 Certificate Bundle feature and not using this feature would save at last this many bytes from the firmware size.

Some of the object files are linked from binary libraries and therefore you will not find a corresponding source file. To locate which component a source file belongs to, it is generally possible to search in the ESP-IDF source tree or look in the Linker Map File for the full path.

Comparing Two Binaries

If making some changes that affect binary size, it is possible to use an ESP-IDF tool to break down the exact differences in size.

This operation is not part of idf.py, it is necessary to run the esp_idf_size Python tool directly.

To do so, first, locate the linker map file with the name PROJECTNAME.map in the build directory. The esp_idf_size tool performs its analysis based on the output of the linker map file.

To compare with another binary, you also need its corresponding .map file saved from the build directory.

For example, to compare two builds, one of which with the default CONFIG_COMPILER_OPTIMIZATION setting Debug (-Og) configuration while another with Optimize for size (-Os):

$ python -m esp_idf_size --diff build_Og/https_request.map build_Os/https_request.map
<CURRENT> MAP file: build_Os/https_request.map
<REFERENCE> MAP file: build_Og/https_request.map
Difference is counted as <CURRENT> - <REFERENCE>, i.e. a positive number means that <CURRENT> is larger.
Total sizes of <CURRENT>:                                                 <REFERENCE>     Difference
 DRAM .data size:   14516 bytes                                                 14956           -440
 DRAM .bss  size:   15792 bytes                                                 15808            -16
Used static DRAM:   30308 bytes ( 150428 available, 16.8% used)                 30764           -456 (   +456 available,      +0 total)
Used static IRAM:   78498 bytes (  52574 available, 59.9% used)                 83918          -5420 (  +5420 available,      +0 total)
      Flash code:  509183 bytes                                                559943         -50760
    Flash rodata:  170592 bytes                                                176736          -6144
Total image size:~ 772789 bytes (.bin may be padded larger)                    835553         -62764

We can see from the Difference column that changing this one setting caused the whole binary to be over 60 KB smaller and over 5 KB more RAM is available.

It is also possible to use the diff mode to output a table of component-level (static library archive) differences:

Note

To get the output in JSON or CSV format using esp_idf_size, it is possible to use the --format option.

python -m esp_idf_size --archives --diff build_Og/https_request.map build_Oshttps_request.map

Also at the individual source file level:

python -m esp_idf_size --files --diff build_Og/https_request.map build_Oshttps_request.map

Other options, like writing the output to a file, are available, pass --help to see the full list.

Showing Size When Linker Fails

If too much static memory is allocated, the linker will fail with an error such as DRAM segment data does not fit, region `iram0_0_seg' overflowed by 44 bytes, or similar.

In these cases, idf.py size will not succeed either. However, it is possible to run esp_idf_size manually to view the partial static memory usage. The memory usage will miss the variables that could not be linked, so there still appears to be some free space.

The map file argument is <projectname>.map in the build directory.

python -m esp_idf_size build/project_name.map

It is also possible to view the equivalent of size-components or size-files output:

python -m esp_idf_size --archives build/project_name.map
python -m esp_idf_size --files build/project_name.map

Linker Map File

Note

This is an advanced analysis method, but it can be very useful. Feel free to skip ahead to Reducing Overall Size and possibly come back to this later.

The idf.py size analysis tools all work by parsing the GNU binutils linker map file, which is a summary of everything the linker did when it created (i.e., linked) the final firmware binary file.

Linker map files themselves are plain text files, so it is possible to read them and find out exactly what the linker did. However, they are also very complex and long, often exceeding 100,000 lines.

The map file itself is broken into parts and each part has a heading. The parts are:

  • Archive member included to satisfy reference by file (symbol)

    • This shows you: for each object file included in the link, what symbol (function or variable) was the linker searching for when it included that object file.

    • If you are wondering why some object file in particular was included in the binary, this part may give a clue. This part can be used in conjunction with the Cross Reference Table at the end of the file.

    Note

    Not every object file shown in this list ends up included in the final binary, some end up in the Discarded input sections list instead.

  • Allocating common symbols

    • This is a list of some global variables along with their sizes. Common symbols have a particular meaning in ELF binary files, but ESP-IDF does not make much use of them.

  • Discarded input sections

    • These sections were read by the linker as part of an object file to be linked into the final binary, but then nothing else referred to them, so they were discarded from the final binary.

    • For ESP-IDF, this list can be very long, as we compile each function and static variable to a unique section in order to minimize the final binary size. Specifically, ESP-IDF uses compiler options -ffunction-sections -fdata-sections and linker option --gc-sections.

    • Items mentioned in this list do not contribute to the final binary.

  • Memory Configuration, Linker script and memory map

    • These two parts go together. Some of the output comes directly from the linker command line and the Linker Script, both provided by Build System. The linker script is partially generated from the ESP-IDF project using the Linker Script Generation feature.

    • As the output of the Linker script and memory map part of the map unfolds, you can see each symbol (function or static variable) linked into the final binary along with its address (as a 16 digit hex number), its length (also in hex), and the library and object file it was linked from (which can be used to determine the component and the source file).

    • Following all of the output sections that take up space in the final .bin file, the memory map also includes some sections in the ELF file that are only used for debugging, e.g., ELF sections .debug_*, etc. These do not contribute to the final binary size. You can notice the address of these symbols is a very small number, starting from 0x0000000000000000 and counting up.

  • Cross Reference Table

    • This table shows the symbol (function or static variable) that the list of object file(s) refers to. If you are wondering why a particular thing is included in the binary, this will help determine what included it.

    Note

    Unfortunately, the Cross Reference Table does not only include symbols that made it into the final binary. It also includes symbols in discarded sections. Therefore, just because something is shown here does not mean that it was included in the final binary - this needs to be checked separately.

Note

Linker map files are generated by the GNU binutils linker ld, not ESP-IDF. You can find additional information online about the linker map file format. This quick summary is written from the perspective of ESP-IDF build system in particular.

Reducing Overall Size

The following configuration options reduces the final binary size of almost any ESP-IDF project:

Note

In addition to the many configuration items shown here, there are a number of configuration options where changing the option from the default increases binary size. These are not noted here. Where the increase is significant is usually noted in the configuration item help text.

Targeted Optimizations

The following binary size optimizations apply to a particular component or a function:

Wi-Fi

lwIP IPv6

  • Setting CONFIG_LWIP_IPV6 to false will reduce the size of the lwIP TCP/IP stack, at the cost of only supporting IPv4.

    Note

    IPv6 is required by some components such as ASIO Port. These components will not be available if IPV6 is disabled.

lwIP IPv4

  • If IPv4 connectivity is not required, setting CONFIG_LWIP_IPV4 to false will reduce the size of the lwIP, supporting IPv6-only TCP/IP stack.

    Note

    Before disabling IPv4 support, please note that IPv6 only network environments are not ubiquitous and must be supported in the local network, e.g., by your internet service provider or using constrained local network settings.

Newlib Nano Formatting

By default, ESP-IDF uses Newlib "full" formatting for I/O functions (printf(), scanf(), etc.)

Enabling the config option CONFIG_NEWLIB_NANO_FORMAT will switch Newlib to the "Nano" formatting mode. This is smaller in code size, and a large part of the implementation is compiled into the ESP32-S2 ROM, so it does not need to be included in the binary at all.

The exact difference in binary size depends on which features the firmware uses, but 25 KB ~ 50 KB is typical.

Enabling "Nano" formatting reduces the stack usage of each function that calls printf() or another string formatting function, see Reducing Stack Sizes.

"Nano" formatting does not support 64-bit integers, or C99 formatting features. For a full list of restrictions, search for --enable-newlib-nano-formatted-io in the Newlib README file.

MbedTLS Features

Under Component Config > mbedTLS, there are multiple mbedTLS features enabled default, some of which can be disabled if not needed to save code size.

These include:

The help text for each option has some more information for reference.

Important

It is strongly not recommended to disable all these mbedTLS options. Only disable options of which you understand the functionality and are certain that it is not needed in the application. In particular:

  • Ensure that any TLS server(s) the device connects to can still be used. If the server is controlled by a third party or a cloud service, it is recommended to ensure that the firmware supports at least two of the supported cipher suites in case one is disabled in a future update.

  • Ensure that any TLS client(s) that connect to the device can still connect with supported/recommended cipher suites. Note that future versions of client operating systems may remove support for some features, so it is recommended to enable multiple supported cipher suites, or algorithms for redundancy.

If depending on third party clients or servers, always pay attention to announcements about future changes to supported TLS features. If not, the ESP32-S2 device may become inaccessible if support changes.

Note

Not every combination of mbedTLS compile-time config is tested in ESP-IDF. If you find a combination that fails to compile or function as expected, please report the details on GitHub.

VFS

Virtual Filesystem Component feature in ESP-IDF allows multiple filesystem drivers and file-like peripheral drivers to be accessed using standard I/O functions (open, read, write, etc.) and C library functions (fopen, fread, fwrite, etc.). When filesystem or file-like peripheral driver functionality is not used in the application, this feature can be fully or partially disabled. VFS component provides the following configuration options:

  • CONFIG_VFS_SUPPORT_TERMIOS — can be disabled if the application does not use termios family of functions. Currently, these functions are implemented only for UART VFS driver. Most applications can disable this option. Disabling this option reduces the code size by about 1.8 KB.

  • CONFIG_VFS_SUPPORT_SELECT — can be disabled if the application does not use the select function with file descriptors. Currently, only the UART and eventfd VFS drivers implement select support. Note that when this option is disabled, select can still be used for socket file descriptors. Disabling this option reduces the code size by about 2.7 KB.

  • CONFIG_VFS_SUPPORT_DIR — can be disabled if the application does not use directory-related functions, such as readdir (see the description of this option for the complete list). Applications that only open, read and write specific files and do not need to enumerate or create directories can disable this option, reducing the code size by 0.5 KB or more, depending on the filesystem drivers in use.

  • CONFIG_VFS_SUPPORT_IO — can be disabled if the application does not use filesystems or file-like peripheral drivers. This disables all VFS functionality, including the three options mentioned above. When this option is disabled, Console can not be used. Note that the application can still use standard I/O functions with socket file descriptors when this option is disabled. Compared to the default configuration, disabling this option reduces code size by about 9.4 KB.

Bootloader Size

This document deals with the size of an ESP-IDF app binary only, and not the ESP-IDF Second Stage Bootloader.

For a discussion of ESP-IDF bootloader binary size, see Bootloader Size.

IRAM Binary Size

If the IRAM section of a binary is too large, this issue can be resolved by reducing IRAM memory usage. See Optimizing IRAM Usage.