Notes on Building Software

Building Software as an Unprivileged (non-root) User

The golden rule of Unix System Administration is to use your superpowers only when necessary. Hence, BLFS recommends that you build software as an unprivileged user and only become the root user when installing the software. This philosophy is followed in all the packages in this book. Unless otherwise specified, all instructions should be executed as an unprivileged user. The book will advise you on instructions that need root privileges.

Unpacking the Software

If a file is in .tar format and compressed, it is unpacked by running one of the following commands:

tar -xvf filename.tar.gz
tar -xvf filename.tgz
tar -xvf filename.tar.Z
tar -xvf filename.tar.bz2

You can also use a slightly different method:

bzcat filename.tar.bz2 | tar -xv

Finally, sometimes we have a compressed patch file in .patch.gz or .patch.bz2 format. The best way to apply the patch is piping the output of the decompressor to the patch utility. For example:

gzip -cd ../patchname.patch.gz | patch -p1

Or for a patch compressed with bzip2:

bzcat ../patchname.patch.bz2 | patch -p1

Verifying File Integrity

Generally, to verify that the downloaded file is complete, many package maintainers also distribute md5sums of the files. To verify the md5sum of the downloaded files, download both the file and the corresponding md5sum file to the same directory (preferably from different on-line locations), and (assuming file.md5sum is the md5sum file downloaded) run the following command:

md5sum -c file.md5sum

If there are any errors, they will be reported. Note that the BLFS book includes md5sums for all the source files also. To use the BLFS supplied md5sums, you can create a file.md5sum (place the md5sum data and the exact name of the downloaded file on the same line of a file, separated by white space) and run the command shown above. Alternately, simply run the command shown below and compare the output to the md5sum data shown in the BLFS book.

md5sum <name_of_downloaded_file>

MD5 is not cryptographically secure, so the md5sums are only provided for detecting unmalicious changes to the file content. For example, an error or truncation introduced during network transfer, or a “stealth” update to the package from the upstream (updating the content of a released tarball instead of making a new release properly).

There is no “100%” secure way to make sure the genuity of the source files. Assuming the upstream is managing their website correctly (the private key is not leaked and the domain is not hijacked), and the trust anchors have been set up correctly using make-ca-1.14 on the BLFS system, we can reasonably trust download URLs to the upstream official website with https protocol. Note that BLFS book itself is published on a website with https, so you should already have some confidence in https protocol or you wouldn't trust the book content.

If the package is downloaded from an unofficial location (for example a local mirror), checksums generated by cryptographically secure digest algorithms (for example SHA256) can be used to verify the genuity of the package. Download the checksum file from the upstream official website (or somewhere you can trust) and compare the checksum of the package from unofficial location with it. For example, SHA256 checksum can be checked with the command:

sha256sum -c file.sha256sum

If GnuPG-2.4.5 is installed, you can also verify the genuity of the package with a GPG signature. Import the upstream GPG public key with:

gpg --recv-key keyID

keyID should be replaced with the key ID from somewhere you can trust (for example, copy it from the upstream official website using https). Now you can verify the signature with:

gpg --recv-key file.sig file

The advantage of GnuPG signature is, once you imported a public key which can be trusted, you can download both the package and its signature from the same unofficial location and verify them with the public key. So you won't need to connect to the official upstream website to retrieve a checksum for each new release. You only need to update the public key if it's expired or revoked.

Creating Log Files During Installation

For larger packages, it is convenient to create log files instead of staring at the screen hoping to catch a particular error or warning. Log files are also useful for debugging and keeping records. The following command allows you to create an installation log. Replace <command> with the command you intend to execute.

( <command> 2>&1 | tee compile.log && exit $PIPESTATUS )

2>&1 redirects error messages to the same location as standard output. The tee command allows viewing of the output while logging the results to a file. The parentheses around the command run the entire command in a subshell and finally the exit $PIPESTATUS command ensures the result of the <command> is returned as the result and not the result of the tee command.

Using Multiple Processors

For many modern systems with multiple processors (or cores) the compilation time for a package can be reduced by performing a "parallel make" by either setting an environment variable or telling the make program to simultaneously execute multiple jobs.

For instance, an Intel Core i9-13900K CPU contains 8 performance (P) cores and 16 efficiency (E) cores, and the P cores support SMT (Simultaneous MultiThreading, also known as “Hyper-Threading”) so each P core can run two threads simultaneously and the Linux kernel will treat each P core as two logical cores. As a result, there are 32 logical cores in total. To utilize all these logical cores running make, we can set an environment variable to tell make to run 32 jobs simultaneously:

export MAKEFLAGS='-j32'

or just building with:

make -j32

If you have applied the optional sed when building ninja in LFS, you can use:

export NINJAJOBS=32

when a package uses ninja, or just:

ninja -j32

If you are not sure about the number of logical cores, run the nproc command.

For make, the default number of jobs is 1. But for ninja, the default number of jobs is N + 2 if the number of logical cores N is greater than 2; or N + 1 if N is 1 or 2. The reason to use a number of jobs slightly greater than the number of logical cores is keeping all logical processors busy even if some jobs are performing I/O operations.

Note that the -j switches only limits the parallel jobs started by make or ninja, but each job may still spawn its own processes or threads. For example, ld.gold will use multiple threads for linking, and some tests of packages can spawn multiple threads for testing thread safety properties. There is no generic way for the building system to know the number of processes or threads spawned by a job. So generally we should not consider the value passed with -j a hard limit of the number of logical cores to use. Read the section called “Use Linux Control Group to Limit the Resource Usage” if you want to set such a hard limit.

Generally the number of processes should not exceed the number of cores supported by the CPU too much. To list the processors on your system, issue: grep processor /proc/cpuinfo.

In some cases, using multiple processes may result in a race condition where the success of the build depends on the order of the commands run by the make program. For instance, if an executable needs File A and File B, attempting to link the program before one of the dependent components is available will result in a failure. This condition usually arises because the upstream developer has not properly designated all the prerequisites needed to accomplish a step in the Makefile.

If this occurs, the best way to proceed is to drop back to a single processor build. Adding -j1 to a make command will override the similar setting in the MAKEFLAGS environment variable.

Use Linux Control Group to Limit the Resource Usage

Sometimes we want to limit the resource usage when we build a package. For example, when we have 8 logical cores, we may want to use only 6 cores for building the package and reserve another 2 cores for playing a movie. The Linux kernel provides a feature called control groups (cgroup) for such a need.

Enable control group in the kernel configuration, then rebuild the kernel and reboot if necessary:

General setup --->
  [*] Control Group support --->                                       [CGROUPS]
    [*] Memory controller                                                [MEMCG]
    [*] Cpuset controller                                              [CPUSETS]

Ensure Systemd-256.5 and Shadow-4.16.0 have been rebuilt with Linux-PAM-1.6.1 support (if you are interacting via a SSH or graphical session, also ensure the OpenSSH-9.9p1 server or the desktop manager has been built with Linux-PAM-1.6.1). As the root user, create a configuration file to allow resource control without root privilege, and instruct systemd to reload the configuration:

mkdir -pv /etc/systemd/system/user@.service.d &&
cat > /etc/systemd/system/user@.service.d/delegate.conf << EOF &&
[Service]
Delegate=memory cpuset
EOF
systemctl daemon-reload

Then logout and login again. Now to run make -j5 with the first 4 logical cores and 8 GB of system memory, issue:

systemctl   --user start dbus                &&
systemd-run --user --pty --pipe --wait -G -d \
            -p MemoryHigh=8G                 \
            -p AllowedCPUs=0-3               \
            make -j5

With MemoryHigh=8G , a soft limit of memory usage is set. If the processes in the cgroup (make and all the descendants of it) uses more than 8 GB of system memory in total, the kernel will throttle down the processes and try to reclaim the system memory from them. But they can still use more than 8 GB of system memory. If you want to make a hard limit instead, replace MemoryHigh with MemoryMax. But doing so will cause the processes killed if 8 GB is not enough for them.

AllowedCPUs=0-3 makes the kernel only run the processes in the cgroup on the logical cores with numbers 0, 1, 2, or 3. You may need to adjust this setting based the mapping between the logical cores and the physical cores. For example, with an Intel Core i9-13900K CPU, the logical cores 0, 2, 4, ..., 14 are mapped to the first threads of the eight physical P cores, the logical cores 1, 3, 5, ..., 15 are mapped to the second threads of the physical P cores, and the logical cores 16, 17, ..., 31 are mapped to the 16 physical E cores. So if we want to use four threads from four different P cores, we need to specify 0,2,4,6 instead of 0-3. Note that the other CPU models may use a different mapping scheme. If you are not sure about the mapping between the logical cores and the physical cores, run the lscpu --extended command which will output logical core IDs in the CPU column, and physical core IDs in the CORE column.

When the nproc or ninja command runs in a cgroup, it will use the number of logical cores assigned to the cgroup as the “system logical core count.” For example, in a cgroup with logical cores 0-3 assigned, nproc will print 4, and ninja will run 6 (4 + 2) jobs simultaneously if no -j setting is explicitly given.

Read the man pages systemd-run(1) and systemd.resource-control(5) for the detailed explanation of parameters in the command.

Automated Building Procedures

There are times when automating the building of a package can come in handy. Everyone has their own reasons for wanting to automate building, and everyone goes about it in their own way. Creating Makefiles, Bash scripts, Perl scripts or simply a list of commands used to cut and paste are just some of the methods you can use to automate building BLFS packages. Detailing how and providing examples of the many ways you can automate the building of packages is beyond the scope of this section. This section will expose you to using file redirection and the yes command to help provide ideas on how to automate your builds.

File Redirection to Automate Input

You will find times throughout your BLFS journey when you will come across a package that has a command prompting you for information. This information might be configuration details, a directory path, or a response to a license agreement. This can present a challenge to automate the building of that package. Occasionally, you will be prompted for different information in a series of questions. One method to automate this type of scenario requires putting the desired responses in a file and using redirection so that the program uses the data in the file as the answers to the questions.

This effectively makes the test suite use the responses in the file as the input to the questions. Occasionally you may end up doing a bit of trial and error determining the exact format of your input file for some things, but once figured out and documented you can use this to automate building the package.

Using yes to Automate Input

Sometimes you will only need to provide one response, or provide the same response to many prompts. For these instances, the yes command works really well. The yes command can be used to provide a response (the same one) to one or more instances of questions. It can be used to simulate pressing just the Enter key, entering the Y key or entering a string of text. Perhaps the easiest way to show its use is in an example.

First, create a short Bash script by entering the following commands:

cat > blfs-yes-test1 << "EOF"
#!/bin/bash

echo -n -e "\n\nPlease type something (or nothing) and press Enter ---> "

read A_STRING

if test "$A_STRING" = ""; then A_STRING="Just the Enter key was pressed"
else A_STRING="You entered '$A_STRING'"
fi

echo -e "\n\n$A_STRING\n\n"
EOF
chmod 755 blfs-yes-test1

Now run the script by issuing ./blfs-yes-test1 from the command line. It will wait for a response, which can be anything (or nothing) followed by the Enter key. After entering something, the result will be echoed to the screen. Now use the yes command to automate the entering of a response:

yes | ./blfs-yes-test1

Notice that piping yes by itself to the script results in y being passed to the script. Now try it with a string of text:

yes 'This is some text' | ./blfs-yes-test1

The exact string was used as the response to the script. Finally, try it using an empty (null) string:

yes '' | ./blfs-yes-test1

Notice this results in passing just the press of the Enter key to the script. This is useful for times when the default answer to the prompt is sufficient. This syntax is used in the Net-tools instructions to accept all the defaults to the many prompts during the configuration step. You may now remove the test script, if desired.

File Redirection to Automate Output

In order to automate the building of some packages, especially those that require you to read a license agreement one page at a time, requires using a method that avoids having to press a key to display each page. Redirecting the output to a file can be used in these instances to assist with the automation. The previous section on this page touched on creating log files of the build output. The redirection method shown there used the tee command to redirect output to a file while also displaying the output to the screen. Here, the output will only be sent to a file.

Again, the easiest way to demonstrate the technique is to show an example. First, issue the command:

ls -l /usr/bin | less

Of course, you'll be required to view the output one page at a time because the less filter was used. Now try the same command, but this time redirect the output to a file. The special file /dev/null can be used instead of the filename shown, but you will have no log file to examine:

ls -l /usr/bin | less > redirect_test.log 2>&1

Notice that this time the command immediately returned to the shell prompt without having to page through the output. You may now remove the log file.

The last example will use the yes command in combination with output redirection to bypass having to page through the output and then provide a y to a prompt. This technique could be used in instances when otherwise you would have to page through the output of a file (such as a license agreement) and then answer the question of do you accept the above?. For this example, another short Bash script is required:

cat > blfs-yes-test2 << "EOF"
#!/bin/bash

ls -l /usr/bin | less

echo -n -e "\n\nDid you enjoy reading this? (y,n) "

read A_STRING

if test "$A_STRING" = "y"; then A_STRING="You entered the 'y' key"
else A_STRING="You did NOT enter the 'y' key"
fi

echo -e "\n\n$A_STRING\n\n"
EOF
chmod 755 blfs-yes-test2

This script can be used to simulate a program that requires you to read a license agreement, then respond appropriately to accept the agreement before the program will install anything. First, run the script without any automation techniques by issuing ./blfs-yes-test2.

Now issue the following command which uses two automation techniques, making it suitable for use in an automated build script:

yes | ./blfs-yes-test2 > blfs-yes-test2.log 2>&1

If desired, issue tail blfs-yes-test2.log to see the end of the paged output, and confirmation that y was passed through to the script. Once satisfied that it works as it should, you may remove the script and log file.

Finally, keep in mind that there are many ways to automate and/or script the build commands. There is not a single “correct” way to do it. Your imagination is the only limit.

Dependencies

For each package described, BLFS lists the known dependencies. These are listed under several headings, whose meaning is as follows:

Using the Most Current Package Sources

On occasion you may run into a situation in the book when a package will not build or work properly. Though the Editors attempt to ensure that every package in the book builds and works properly, sometimes a package has been overlooked or was not tested with this particular version of BLFS.

If you discover that a package will not build or work properly, you should see if there is a more current version of the package. Typically this means you go to the maintainer's web site and download the most current tarball and attempt to build the package. If you cannot determine the maintainer's web site by looking at the download URLs, use Google and query the package's name. For example, in the Google search bar type: 'package_name download' (omit the quotes) or something similar. Sometimes typing: 'package_name home page' will result in you finding the maintainer's web site.

Stripping One More Time

In LFS, stripping of debugging symbols and unneeded symbol table entries was discussed a couple of times. When building BLFS packages, there are generally no special instructions that discuss stripping again. Stripping can be done while installing a package, or afterwards.

Stripping while Installing a Package

There are several ways to strip executables installed by a package. They depend on the build system used (see below the section about build systems), so only some generalities can be listed here:

Stripping Installed Executables

The strip utility changes files in place, which may break anything using it if it is loaded in memory. Note that if a file is in use but just removed from the disk (i.e. not overwritten nor modified), this is not a problem since the kernel can use “deleted” files. Look at /proc/*/maps and it is likely that you'll see some (deleted) entries. The mv just removes the destination file from the directory but does not touch its content, so that it satisfies the condition for the kernel to use the old (deleted) file. But this approach can detach hard links into duplicated copies, causing a bloat which is obviously unwanted as we are stripping to reduce system size. If two files in a same file system share the same inode number, they are hard links to each other and we should reconstruct the link. The script below is just an example. It should be run as the root user:

cat > /usr/sbin/strip-all.sh << "EOF"
#!/usr/bin/bash

if [ $EUID -ne 0 ]; then
  echo "Need to be root"
  exit 1
fi

last_fs_inode=
last_file=

{ find /usr/lib -type f -name '*.so*' ! -name '*dbg'
  find /usr/lib -type f -name '*.a'
  find /usr/{bin,sbin,libexec} -type f
} | xargs stat -c '%m %i %n' | sort | while read fs inode file; do
       if ! readelf -h $file >/dev/null 2>&1; then continue; fi
       if file $file | grep --quiet --invert-match 'not stripped'; then continue; fi

       if [ "$fs $inode" = "$last_fs_inode" ]; then
         ln -f $last_file $file;
         continue;
       fi

       cp --preserve $file    ${file}.tmp
       strip --strip-unneeded ${file}.tmp
       mv ${file}.tmp $file

       last_fs_inode="$fs $inode"
       last_file=$file
done
EOF
chmod 744 /usr/sbin/strip-all.sh

If you install programs in other directories such as /opt or /usr/local, you may want to strip the files there too. Just add other directories to scan in the compound list of find commands between the braces.

For more information on stripping, see https://www.technovelty.org/linux/stripping-shared-libraries.html.

Working with different build systems

There are now three different build systems in common use for converting C or C++ source code into compiled programs or libraries and their details (particularly, finding out about available options and their default values) differ. It may be easiest to understand the issues caused by some choices (typically slow execution or unexpected use of, or omission of, optimizations) by starting with the CFLAGS, CXXFLAGS, and LDFLAGS environment variables. There are also some programs which use Rust.

Most LFS and BLFS builders are probably aware of the basics of CFLAGS and CXXFLAGS for altering how a program is compiled. Typically, some form of optimization is used by upstream developers (-O2 or -O3), sometimes with the creation of debug symbols (-g), as defaults.

If there are contradictory flags (e.g. multiple different -O values), the last value will be used. Sometimes this means that flags specified in environment variables will be picked up before values hardcoded in the Makefile, and therefore ignored. For example, where a user specifies -O2 and that is followed by -O3 the build will use -O3.

There are various other things which can be passed in CFLAGS or CXXFLAGS, such as allowing using the instruction set extensions available with a specific microarchitecture (e.g. -march=amdfam10 or -march=native), tune the generated code for a specific microarchitecture (e. g. -mtune=tigerlake or -mtune=native, if -mtune= is not used, the microarchitecture from -march= setting will be used), or specifying a specific standard for C or C++ (-std=c++17 for example). But one thing which has now come to light is that programmers might include debug assertions in their code, expecting them to be disabled in releases by using -D NDEBUG. Specifically, if Mesa-24.2.5 is built with these assertions enabled, some activities such as loading levels of games can take extremely long times, even on high-class video cards.

Autotools with Make

This combination is often described as “CMMI” (configure, make, make install) and is used here to also cover the few packages which have a configure script that is not generated by autotools.

Sometimes running ./configure --help will produce useful options about switches which might be used. At other times, after looking at the output from configure you may need to look at the details of the script to find out what it was actually searching for.

Many configure scripts will pick up any CFLAGS or CXXFLAGS from the environment, but CMMI packages vary about how these will be mixed with any flags which would otherwise be used (variously: ignored, used to replace the programmer's suggestion, used before the programmer's suggestion, or used after the programmer's suggestion).

In most CMMI packages, running make will list each command and run it, interspersed with any warnings. But some packages try to be “silent” and only show which file they are compiling or linking instead of showing the command line. If you need to inspect the command, either because of an error, or just to see what options and flags are being used, adding V=1 to the make invocation may help.

CMake

CMake works in a very different way, and it has two backends which can be used on BLFS: make and ninja. The default backend is make, but ninja can be faster on large packages with multiple processors. To use ninja, specify -G Ninja in the cmake command. However, there are some packages which create fatal errors in their ninja files but build successfully using the default of Unix Makefiles.

The hardest part of using CMake is knowing what options you might wish to specify. The only way to get a list of what the package knows about is to run cmake -LAH and look at the output for that default configuration.

Perhaps the most-important thing about CMake is that it has a variety of CMAKE_BUILD_TYPE values, and these affect the flags. The default is that this is not set and no flags are generated. Any CFLAGS or CXXFLAGS in the environment will be used. If the programmer has coded any debug assertions, those will be enabled unless -D NDEBUG is used. The following CMAKE_BUILD_TYPE values will generate the flags shown, and these will come after any flags in the environment and therefore take precedence.

CMake tries to produce quiet builds. To see the details of the commands which are being run, use make VERBOSE=1 or ninja -v.

By default, CMake treats file installation differently from the other build systems: if a file already exists and is not newer than a file that would overwrite it, then the file is not installed. This may be a problem if a user wants to record which file belongs to a package, either using LD_PRELOAD, or by listing files newer than a timestamp. The default can be changed by setting the variable CMAKE_INSTALL_ALWAYS to 1 in the environment, for example by export'ing it.

Meson

Meson has some similarities to CMake, but many differences. To get details of the defines that you may wish to change you can look at meson_options.txt which is usually in the top-level directory.

If you have already configured the package by running meson and now wish to change one or more settings, you can either remove the build directory, recreate it, and use the altered options, or within the build directory run meson configure, e.g. to set an option:

meson configure -D <some_option>=true

If you do that, the file meson-private/cmd_line.txt will show the last commands which were used.

Meson provides the following buildtype values, and the flags they enable come after any flags supplied in the environment and therefore take precedence.

The -D NDEBUG flag is implied by the release buildtype for some packages (for example Mesa-24.2.5). It can also be provided explicitly by passing -D b_ndebug=true.

To see the details of the commands which are being run in a package using meson, use ninja -v.

Rustc and Cargo

Most released rustc programs are provided as crates (source tarballs) which will query a server to check current versions of dependencies and then download them as necessary. These packages are built using cargo --release. In theory, you can manipulate the RUSTFLAGS to change the optimize-level (default for --release is 3, i. e. -Copt-level=3, like -O3) or to force it to build for the machine it is being compiled on, using -Ctarget-cpu=native but in practice this seems to make no significant difference.

If you are compiling a standalone Rust program (as an unpackaged .rs file) by running rustc directly, you should specify -O (the abbreviation of -Copt-level=2) or -Copt-level=3 otherwise it will do an unoptimized compile and run much slower. If you are compiling the program for debugging it, replace the -O or -Copt-level= options with -g to produce an unoptimized program with debug info.

Like ninja, by default cargo uses all logical cores. This can often be worked around, either by exporting CARGO_BUILD_JOBS=<N> or passing --jobs <N> to cargo. For compiling rustc itself, specifying --jobs <N> for invocations of x.py (together with the CARGO_BUILD_JOBS environment variable, which looks like a “belt and braces” approach but seems to be necessary) mostly works. The exception is running the tests when building rustc, some of them will nevertheless use all online CPUs, at least as of rustc-1.42.0.

Optimizing the build

Many people will prefer to optimize compiles as they see fit, by providing CFLAGS or CXXFLAGS. For an introduction to the options available with gcc and g++ see https://gcc.gnu.org/onlinedocs/gcc-14.2.0/gcc/Optimize-Options.html. The same content can be also found in info gcc.

Some packages default to -O2 -g, others to -O3 -g, and if CFLAGS or CXXFLAGS are supplied they might be added to the package's defaults, replace the package's defaults, or even be ignored. There are details on some desktop packages which were mostly current in April 2019 at https://www.linuxfromscratch.org/~ken/tuning/ - in particular, README.txt, tuning-1-packages-and-notes.txt, and tuning-notes-2B.txt. The particular thing to remember is that if you want to try some of the more interesting flags you may need to force verbose builds to confirm what is being used.

Clearly, if you are optimizing your own program you can spend time to profile it and perhaps recode some of it if it is too slow. But for building a whole system that approach is impractical. In general, -O3 usually produces faster programs than -O2. Specifying -march=native is also beneficial, but means that you cannot move the binaries to an incompatible machine - this can also apply to newer machines, not just to older machines. For example programs compiled for amdfam10 run on old Phenoms, Kaveris, and Ryzens, but programs compiled for a Kaveri will not run on a Ryzen because certain op-codes are not present. Similarly, if you build for a Haswell not everything will run on a SandyBridge.

When a shared library is built by GCC, a feature named “semantic interposition” is enabled by default. When the shared library refers to a symbol name with external linkage and default visibility, if the symbol exists in both the shared library and the main executable, semantic interposition guarantees the symbol in the main executable is always used. This feature was invented in an attempt to make the behavior of linking a shared library and linking a static library as similar as possible. Today only a small number of packages still depend on semantic interposition, but the feature is still on by the default of GCC, causing many optimizations disabled for shared libraries because they conflict with semantic interposition. The -fno-semantic-interposition option can be passed to gcc or g++ to disable semantic interposition and enable more optimizations for shared libraries. This option is used as the default of some packages (for example Python-3.12.7), and it's also the default of Clang.

There are also various other options which some people claim are beneficial. At worst, you get to recompile and test, and then discover that in your usage the options do not provide a benefit.

If building Perl or Python modules, in general the CFLAGS and CXXFLAGS used are those which were used by those “parent” packages.

For LDFLAGS, there are three options can be used for optimization. They are quite safe to use and the building system of some packages use some of these options as the default.

With -Wl,-O1, the linker will optimize the hash table to speed up the dynamic linking. Note that -Wl,-O1 is completely unrelated to the compiler optimization flag -O1.

With -Wl,--as-needed, the linker will disregard unnecessary -lfoo options from the command line, i. e. the shared library libfoo will only be linked if a symbol in libfoo is really referred from the executable or shared library being linked. This can sometimes mitigate the “excessive dependencies to shared libraries” issues caused by libtool.

With -Wl,-z,pack-relative-relocs, the linker generates a more compacted form of the relative relocation entries for PIEs and shared libraries. It reduces the size of the linked PIE or shared library, and speeds up the loading of the PIE or shared library.

The -Wl, prefix is necessary because despite the variable is named LDFLAGS, its content is actually passed to gcc (or g++, clang, etc.) during the link stage, not directly passed to ld.

Options for hardening the build

Even on desktop systems, there are still a lot of exploitable vulnerabilities. For many of these, the attack comes via javascript in a browser. Often, a series of vulnerabilities are used to gain access to data (or sometimes to pwn, i.e. own, the machine and install rootkits). Most commercial distros will apply various hardening measures.

In the past, there was Hardened LFS where gcc (a much older version) was forced to use hardening (with options to turn some of it off on a per-package basis). The current LFS and BLFS books are carrying forward a part of its spirit by enabling PIE (-fPIE -pie) and SSP (-fstack-protector-strong) as the defaults for GCC and clang. And, the linkers (ld.bfd and ld.gold) have also enabled -Wl,-z,relro which makes a part of the Global Offset Table (GOT) immutable, by default since Binutils 2.27. What is being covered here is different - first you have to make sure that the package is indeed using your added flags and not over-riding them.

For hardening options which are reasonably cheap, there is some discussion in the 'tuning' link above (occasionally, one or more of these options might be inappropriate for a package). These options are -D _FORTIFY_SOURCE=2 (or -D _FORTIFY_SOURCE=3 which is more secure but with a larger performance overhead) and (for C++) -D _GLIBCXX_ASSERTIONS. On modern machines these should only have a little impact on how fast things run, and often they will not be noticeable.

The main distros use much more, such as:

In GCC 14, the option -fhardened is a shorthand to enable all the hardening options mentioned above. It sets -D _FORTIFY_SOURCE=3 instead of -D _FORTIFY_SOURCE=2.

You may also encounter the so-called “userspace retpoline” (-mindirect-branch=thunk etc.) which is the equivalent of the spectre mitigations applied to the linux kernel in late 2018. The kernel mitigations caused a lot of complaints about lost performance, if you have a production server you might wish to consider testing that, along with the other available options, to see if performance is still sufficient.

Whilst gcc has many hardening options, clang/LLVM's strengths lie elsewhere. Some options which gcc provides are said to be less effective in clang/LLVM.