GCC optimization
This guide provides an introduction to optimizing compiled code using safe, sane CFLAGS and CXXFLAGS. It also describes the theory behind optimizing in general.
Default CFLAGS can be set in make.conf for Gentoo systems. CFLAGS can also be specified per-package.
For more information see CFLAGS and CXXFLAGS in the Gentoo Handbook, and the safe CFLAGS article. See also the FAQ.
Introduction
What are CFLAGS and CXXFLAGS?
CFLAGS and CXXFLAGS are among the environment variables conventionally used to specify compiler options to a build system when compiling C and C++ code. While these variables are not standardized, their use is essentially ubiquitous and any correctly written build should understand these for passing extra or custom options when it invokes the compiler. See the GNU make info page for a list of some of the commonly used variables in this category.
Because such a large proportion of the packages that make up most Gentoo systems are written in C and C++, these are two variables administrators will definitely want to set correctly as they will greatly influence the way much of the system is built.
They can be used to decrease the amount of debug messages for a program, increase error warning levels and, of course, to optimize the code produced. The GCC manual maintains a complete list of available options and their purposes.
How are they used?
Normally, CFLAGS and CXXFLAGS would be set in the environment when invoking a configure script or with makefiles generated by the automake program. In Gentoo-based systems, set the CFLAGS and CXXFLAGS variables in /etc/portage/make.conf. Variables set in this file will be exported to the environment of programs invoked by portage such that all packages will be compiled using these options as a base.
CFLAGS="-march=skylake -O2 -pipe"
CXXFLAGS="${CFLAGS}"
While it is possible to have multiple lines in USE flags, having multiple lines in CFLAGS can and will result in problems with programs such as cmake. Make sure the CFLAGS declaration is on a single line, with as little whitespace as possible to avoid issues. See bug #500034 as an example.
As seen in the example above, the CXXFLAGS variable is set to use all the options present in CFLAGS. Almost every system should be configured in this manner. Additional options for CXXFLAGS are less common and don't usually apply generally enough to deserve setting them globally.
Safe CFLAGS article might help beginners start optimizing their systems.
Misconceptions
While compiler optimizations enabled by various CFLAGS can be an effective means of producing smaller and/or faster binaries, they can also impair the function of the code, bloat its size, slow down its execution time, or simply cause a build failure. The point of diminishing performance returns is reached rather quickly when dealing with CFLAGS. Don't set them arbitrarily.
Remember, the global CFLAGS configured in /etc/portage/make.conf will be applied to every package on the system so administrators typically only set general, widely-applicable options. Individual packages further modify these options either in the ebuild or the build system itself to generate the final set of flags used when invoking the compiler.
Ready?
Being aware of the risks involved, take a look at some sane, safe optimizations. These will hold in good stead and will be endearing to developers the next time a problem is reported on Bugzilla. (Developers will usually request the user to recompile a package with minimal CFLAGS to see if the problem persists. Remember: aggressive flags can ruin code!)
Optimizing
The basics
The goal behind CFLAGS and CXXFLAGS is to create code tailor-made to the system; it should function perfectly while being lean and fast, if possible. Sometimes these conditions are mutually exclusive, so this guide will stick to combinations known to work well. Ideally, they are the best available for any CPU architecture. For informational purposes, aggressive flag use will be covered later. Not every option listed on the GCC manual (there are hundreds) will be discussed, but basic, most common flags will be reviewed.
When unaware of what a flag does, refer to the relevant chapter of the GCC manual. If still stumped after viewing the manual, try a search engine or check out the GCC mailing lists.
-march
The first and most important option is -march
. This tells the compiler what code it should produce for the system's processor architecture (or arch); it tells GCC that it should produce code for a certain kind of CPU. Different CPUs have different capabilities, support different instruction sets, and have different ways of executing code. The -march
flag will instruct the compiler to produce specific code for the system's CPU, with all its capabilities, features, instruction sets, quirks, and so on provided the source code is prepared to use them. For instance, to take benefit from AVX instructions, the source code needs to be adapted to support it.
-march=
is an ISA selection option; it tells the compiler that it may use the instructions from the ISA. On an Intel/AMD64 platform with -march=native -O2
or lower optimization level, the code will likely end up with AVX instructions used but using shorter SSE XMM registers. When using a compiler version prior to GCC 12, in order to take full advantage of AVX YMM registers, the -ftree-vectorize
or -O3
options should be used as well[1].
Even though the CHOST variable in /etc/portage/make.conf specifies the general architecture used, -march
should still be used so that programs can be optimized for the system specific processor. x86 and x86-64 CPUs (among others) should make use of the -march
flag.
What kind of CPU does the system have? To find out, run the following command:
user $
cat /proc/cpuinfo
or even install app-portage/cpuid2cpuflags and add the available CPU-specific options to the /etc/portage/package.use/00cpuflags file, which the tool does through e.g. the CPU_FLAGS_* variable:
user $
cpuid2cpuflags
CPU_FLAGS_X86: aes avx avx2 f16c fma3 mmx mmxext pclmul popcnt sha sse sse2 sse3 sse4_1 sse4_2 sse4a ssse3
root #
echo "*/* $(cpuid2cpuflags)" > /etc/portage/package.use/00cpu-flags
To get more details, including march
and mtune
values, two commands can be used.
- The first command tells the compiler not to do any linking (
-c
), and instead of interpreting the--help
option for clarifying command line options, it now shows if certain options are enabled or disabled (-Q
). In this case, the options shown are those enabled for the selected target:user $
gcc -c -Q -march=native --help=target
- The second command will show the compiler directives for building the header file, but without actually performing the steps and instead showing them on the screen (
-###
). The final output line is the command that holds all the optimization options and architecture selection:user $
gcc -### -march=native /usr/include/stdlib.h
cpu-type
- The glibc-hwcaps feature (>=sys-libs/glibc-2.33) can be used to define
-march
for a more general processor architecture (for >=sys-devel/gcc-11):
user $
/lib64/ld-linux-x86-64.so.2 --help
... Subdirectories of glibc-hwcaps directories, in priority order: x86-64-v4 x86-64-v3 (supported, searched) x86-64-v2 (supported, searched) x86_64 (supported, searched)
In this example, the cpu supports x86-64-v3 psABI x86_64 which can be used for -march=x86-64-v3
.
Now lets see -march
in action. This example is for an older AMD Athlon 64 chip:
CFLAGS="-march=athlon64"
CXXFLAGS="${CFLAGS}"
Here's another one for a common Intel processor:
CFLAGS="-march=skylake"
CXXFLAGS="${CFLAGS}"
If the type of CPU is undetermined, or if the user does not know what setting to choose, it is possible use the -march=native
setting. When this flag is used, GCC will attempt to detect the processor and automatically set appropriate flags for it. However, this should not be used when intending to compile packages for different CPUs!
If compiling packages on one computer in order to run them on a different computer (such as when using a fast computer to build for an older, slower machine), then do not use -march=native
. "Native" means that the code produced will run only on that type of CPU. The applications built with -march=native
on an Intel Core CPU will not be able to run on an old Intel Atom CPU.
Also available are the -mtune
and -mcpu
flags. These flags are normally only used when there is no available -march
option; certain processor architectures may require -mtune
or even -mcpu
. Unfortunately, GCC's behavior isn't very consistent with how each flag behaves from one architecture to the next.
On x86 and x86-64 CPUs, -march
will generate code specifically for that CPU using its available instruction sets and the correct ABI; it will have no backwards compatibility for older/different CPUs. Consider using -mtune
when generating code for older CPUs such as i386 and i486. -mtune
produces more generic code than -march
; though it will tune code for a certain CPU, it does not take into account available instruction sets and ABI. Do not use -mcpu
on x86 or x86-64 systems, as it is deprecated for those architectures.
Only non-x86/x86-64 CPUs (such as ARM, SPARC, Alpha, and PowerPC) may require -mtune
or -mcpu
instead of -march
. On these architectures, -mtune
/ -mcpu
will sometimes behave just like -march
(on x86/x86-64) but with a different flag name. Again, GCC's behavior and flag naming is not consistent across architectures, so be sure to check the GCC manual to determine which one should be used.
For more suggested
-march
/ -mtune
/ -mcpu
settings, please read chapter 5 of the appropriate Gentoo Installation Handbook for the architecture. Also, read the GCC manual's list of architecture-specific options, as well as more detailed explanations about the differences between -march
, -mcpu
, and -mtune
.Do not use
-march=native
or -mtune=native
in the CFLAGS or CXXFLAGS variables of make.conf when compiling with distcc.-O
Using
-O3
may cause some packages to break either during the compilation or misbehave at runtime, although -O3
retains standard conformance, hence any breakage is either undefined behaviour in the application, or (very rarely) a compiler bug.To print all packages that were built with specified CFLAGS/CXXFLAGS it's possible to use the following command:
grep O3 /var/db/pkg/*/*/CFLAGS
Next up is the -O
variable. This variable controls the overall level of optimization. Changing this value will make the code compilation take more time and will use much more memory, especially as the level of optimization is increased.
There are seven -O
settings: -O0
, -O1
, -O2
, -O3
, -Os
, -Oz
, and -Og
. Only use one of them in /etc/portage/make.conf. There is an eighth option which must not be used system-wide: -Ofast
- despite its name, it breaks standards conformance and is dangerous.
With the exception of -O0
, the -O
settings each activate several additional flags, so be sure to read the GCC manual's chapter on optimization options to learn which flags are activated at each -O
level, as well as some explanations as to what they do.
Let us examine each optimization level:
-O0
: This level (that is the letter "O" followed by a zero) turns off optimization entirely and is the default if no-O
level is specified in CFLAGS or CXXFLAGS. This reduces compilation time and can improve debugging info, but some applications will not work properly without optimization enabled. This option is not recommended except for debugging purposes.
-O1
: the most basic optimization level. The compiler will try to produce faster, smaller code without taking much compilation time. It is basic, but it should get the job done all the time.
-O2
: A step up from-O1
. The recommended level of optimization unless the system has special needs.-O2
will activate a few more flags in addition to the ones activated by-O1
. With-O2
, the compiler will attempt to increase code performance without compromising on size, and without taking too much compilation time.
-O3
: enables-O2
as well as optimizations that are expensive in terms of compile time and memory usage. Compiling with-O3
is likely to improve performance (although not guaranteed). While historically-O3
could expose compiler bugs, nowadays issues with-O3
are almost always instances of undefined behavior (UB) and the code is at fault and should be fixed. Some packages are still known-broken with-O3
. Using-O3
is not recommended unless running test suites for packages.
-Ofast
: New in GCC 4.7, consists of-O3
plus-ffast-math
,-fno-protect-parens
,-fallow-store-data-races
,-fstack-arrays
and-fno-semantic-interposition
. This option breaks strict standards compliance, and is not recommended for use. Do not use this system-wide ever, and only use it per-package if the software has been audited for use with the option.
-Os
: optimizes code for size. It activates all-O2
options that do not increase the size of the generated code. It can be useful for machines that have extremely limited disk storage space and/or CPUs with small cache sizes.
-Oz
: Introduced in GCC 12.1, more aggressively optimize for size than-Os
. Note this will heavily degrade runtime performance than-O2
, due to increasing the number of instructions executed if those instructions require fewer bytes to encode.
-Og
: In GCC 4.8, a new general optimization level,-Og
, has been introduced. It addresses the need for fast compilation and a superior debugging experience while providing a reasonable level of runtime performance. Overall experience for development should be better than the default optimization level-O0
. Note that-Og
does not imply-g
, it simply disables optimizations that may interfere with debugging.
As previously mentioned, -O2
is the recommended optimization level. If package compilation fails and while not using -O2
, try rebuilding with that option. As a fallback option, try setting the CFLAGS and CXXFLAGS to a lower optimization level, such as -O1
or even -O0 -g2 -ggdb
(for error reporting and checking for possible problems).
-pipe
A common flag is -pipe
. This flag has no effect on the generated code, but it makes the compilation process faster. It tells the compiler to use pipes instead of temporary files during the different stages of compilation, which uses more memory. On systems with low memory, GCC might get killed. In those cases do not use this flag.
-ftree-vectorize
-ftree-vectorize
is an optimization option (default at -O2
and up), which attempts to vectorize loops using the selected ISA if possible. The reason it previously wasn't enabled at -O2
is that it doesn't always improve code, it can make code slower as well, and usually makes the code larger; it really depends on the loop etc. As of GCC 12, it is enabled by default at -O2
with a low cost model (-fvect-cost-model=very-cheap
) to strike a balance between code size and speed benefits. The cost model can be specified with -fvect-cost-model
. Alternative vectorization cost models include: cheap
, dynamic
, and unlimited
. The dynamic
cost model (default at -O3
) is going to estimate the cost of the loop using scalar instructions and vector instructions and will be able to decide whether vectorization is profitable using either compile time or runtime checks. The cheap
cost model is similar to dynamic
, although slighty more conservative, it will not take effect if the runtime checks for data dependencies or alignment exceed the parameters. The unlimited
cost model assumes that vectorization is always profitable, making it switch from auto-vectorization to explicit vectorization, but it should never be used system-wide as it will cause severe code bloat.
-fomit-frame-pointer
This is a very common flag designed to reduce generated code size. It is turned on at all levels of -O
(except -O0
) on architectures where doing so does not interfere with debugging (such as x86-64), but it may need to be activated. In that case add it to the flags. Though the GCC manual does not specify all architectures, it is turned on by using the -O
option. It's still necessary to explicitly enable the -fomit-frame-pointer
option to activate it on x86-32 with GCC up to version 4.6, or when using -Os
on x86-32 with any version of GCC. However, using -fomit-frame-pointer
will make debugging hard or impossible. It is worth noting that this option is not enabled by default when using Clang. It is also worth noting that keeping frame pointers is actually beneficial when profiling a code base, and in this case users may want to disable this through -fno-omit-frame-pointer
.
Do not combine
-fomit-frame-pointer
with the similar flag -momit-leaf-frame-pointer
. Using the latter flag is discouraged, as -fomit-frame-pointer
already does the job properly. Furthermore, -momit-leaf-frame-pointer
has been shown to negatively impact code performance.-fno-semantic-interposition
This flag is meant to improve code generation quality and can provide greater performance (default at -Ofast
). The default behavior (-fsemantic-interposition
) follows the ELF standard, which allows interposing of symbols by the dynamic linker. While this might sound useful in certain cases it prevents the compiler from doing extensive code analyses and optimizations (in particular the compiler will not attempt inlining unless the functions have been specifically declared as inlined). Contrary to popular belief, enabling this flag globally is safe (unless interposing symbols is required, for example when using different allocators on system libraries), but the reason for it not being enabled by default is to comply with the ELF standard. In contrast, this flag is part of the default when using Clang.
-msse, -msse2, -msse3, -mmmx, -m3dnow
The x86 and x86-64 architectures can have Streaming SIMD Extensions (SSE), SSE2, SSE3, MMX, and 3DNow! instruction sets. These are useful primarily in multimedia, gaming, and other floating point-intensive computing tasks, though they also contain several other mathematical enhancements. These instruction sets are found in more modern CPUs.
Be sure to see if the CPU supports these instruction sets by running cat /proc/cpuinfo. The output will include any supported additional instruction sets. Note that pni is just a different name for SSE3.
Normally none of these flags need to be added to /etc/portage/make.conf, as long as the system is using the correct -march
(for example, -march=nocona
implies -msse3
). Some notable exceptions are newer VIA and AMD64 CPUs that support instructions not implied by -march
(such as SSE3). For CPUs like these additional flags will need to be enabled where appropriate after checking /proc/cpuinfo.
Graphite
Graphite analyzes loop costs for graphite
GCC when COMMON_FLAGS
has -floop-block -fgraphite-identity -floop-parallelize-all
. Refer to the Optimize Options for more.
Hardening optimizations
While it is possible to use a hardened profile, it certainly isn't necessary for adding some hardening flags to /etc/portage/make.conf on a different profile. Especially on a desktop system, the use of position independent code (PIC) and position independent executables (PIE) on a system-wide level may cause compilation failures.
Reading section Do I need to pass any flags to LDFLAGS/CFLAGS in order to turn on hardened building? in the Hardened/FAQ is recommended for retrieving some basic hardened CFLAGS/CXXFLAGS.
Changing the CFLAGS/CXXFLAGS can cause problems and in some cases may even render a system unusable. Rebuilding the whole system with emerge -e @world may resolve the situation.
Hardening an otherwise unhardened system, like when using a desktop profile, can be considered a GCC optimization as well, especially in the light of security vulnerabilities such as Meltdown and Spectre.
Some packages feature an individual hardened
USE flag, enabling tested security enhancements (like CFLAGS/CXXFLAGS). It may be a good idea to set this system-wide in /etc/portage/make.conf.
Optimizing CFLAGS/CXXFLAGS for overflow protection can be a good idea if security concerns outweigh speed optimization. This may be the case on a daily-use desktop system, while e.g. on an optimized gaming PC it will be considered counterproductive.
Another measure is Address Space Layout Randomization (ASLR) which can increase security by randomly placing each function and buffer in memory. This makes it harder for attack vectors to succeed.
PIE is enabled by default when it is safe to do so on in 17.0 profiles and newer[3]. PIC may also be enabled by default on executables for architectures that require it (like AMD64). There is no need to set PIE or PIC manually in CFLAGS.
A lot of these flags are now applied internally through the GCC toolchain automatically under the hardened profile, and some even under the regular profile. See table at Hardened/Toolchain.
The table below is provided to document flags rather than to serve as a list to use, as defaults are handled elsewhere in Gentoo, and specifying these system-wide may not work correctly:
CFLAGS/CXXFLAGS | LDFLAGS | function |
---|---|---|
-D_FORTIFY_SOURCE=2 (or -D_FORTIFY_SOURCE=3 )
|
run-time buffer overflow detection | |
-D_GLIBCXX_ASSERTIONS
|
run-time bounds checking for C++ strings and containers | |
-fstack-protector-strong
|
stack smashing protector | |
-fstack-clash-protection
|
increased reliability of stack overflow detection | |
-fcf-protection
|
control flow integrity protection | |
-Wl,-z,now
|
disable lazy binding | |
-Wl,-z,relro
|
read-only segments after relocation | |
-fpie
|
-Wl,-pie
|
full ASLR for executables |
-fpic -shared
|
no text relocations for shared libraries |
Optimization FAQs
Higher version of GCC should mean better optimizations?
Not always because of software regression, where an optimization with an earlier version of GCC no longer optimizes. A full list of regressions can be found at this link. Should this happen, please file a bug to Gentoo's bugzilla and/or GCC's bugzilla.
Is there a perfect optimizer/CFLAGS?
No, because it would solve the halting problem, where it can tell if any program stops or runs forever [4]. This means that there is no perfect CFLAGS for every package. Even if such CFLAGS were found, packages change over time.
What about optimizing GCC itself?
gcc has pgo
and lto
use flags that enables Profile Guided Optimization and Link Time Optimization respectively. To enable for building gcc itself with PGO and LTO:
sys-devel/gcc pgo lto
In Gentoo, a 3-stage bootstrap of gcc is done, meaning it compiles itself three times [5]. In stage1, gcc is complied using an older gcc. In stage2, gcc is compiled using stage1 gcc. In stage3, gcc is compiled using stage2 gcc and is used to verify that stage2 gcc and stage3 gcc are the same. This is done because it is tested more completely and has better performance. The lto
use flag adds -flto to BOOT_CFLAGS. The pgo
use flag adds -fprofile-generate
to stage2 gcc and adds -fprofile-use -fprofile-reproducible=parallel-runs
to stage4 gcc.
gcc performance may improve via PGO, although it may as much as double the compile times.
But I get better performance with -funroll-loops -fomg-optimize!
No, people only think they do because someone has convinced them that more flags are better. Aggressive flags will only hurt applications when used system-wide. Even the GCC manual says that using -funroll-loops
and -funroll-all-loops
will make code larger and may run more slowly. Yet for some reason, these two flags, along with -ffast-math
, -fforce-mem
, -fforce-addr
, and similar flags, continue to be very popular among ricers who want the biggest bragging rights. Supposing that these flags are advantageous for a single application, generalizing this to the whole system is unwise.
The truth of the matter is that they are dangerously aggressive flags. Take a good look around the Gentoo Forums and Bugzilla to see what those flags do: nothing good!
These flags are not needed globally in CFLAGS or CXXFLAGS. They will only hurt performance. They might bring on the idea of having a high-performance system running on the bleeding edge, but they don't do anything but bloat the code and get bugs marked INVALID or WONTFIX.
Dangerous flags like these are not needed. Don't use them. Stick to the basics: -march
, -O
, and -pipe
.
What about -O levels higher than 3?
Some users boast about even better performance obtained by using -O4
, -O9
, and so on, but the reality is that -O
levels higher than 3 have no effect. The compiler may accept CFLAGS like -O4
, but it actually doesn't do anything with them. It only performs the optimizations for -O3
, nothing more.
Need more proof? Examine the source code:
case OPT_LEVELS_3_PLUS:
enabled = (level >= 3);
break;
case OPT_LEVELS_3_PLUS_AND_SIZE:
enabled = (level >= 3 || size);
break;
As can be seen, any value higher than 3 is treated as just -O3
.
What about compiling outside the target machine?
Some readers might wonder if compiling outside the target machine with a strictly inferior CPU or GCC sub-architecture will result in inferior optimization results (compared to a native compilation). The answer is simple: No. Regardless of the actual hardware on which the compilation takes place and the CHOST for which GCC was built, as long as the same arguments are used (except for -march=native
) and the same version of GCC is used (although minor version might be different), the resulting optimizations are strictly the same.
To exemplify, if Gentoo is installed on a machine whose GCC's CHOST is i686-pc-linux-gnu, and a Distcc server is setup on another computer whose GCC's CHOST is i486-linux-gnu, then there is no need to be afraid that the results would be less optimal because of the strictly inferior sub-architecture of the remote compiler and/or hardware. The result would be as optimized as a native build, as long as the same options are passed to both compilers (and the -march
parameter doesn't get a native
argument). In this particular case the target architecture needs to be specified explicitly as explained in Distcc.
The only difference in behavior between two GCC versions built targeting different sub-architectures is the implicit default argument for the -march
parameter, which is derived from the GCC's CHOST when not explicitly provided in the command line.
What about redundant flags?
Oftentimes CFLAGS and CXXFLAGS that are turned on at various -O
levels are specified redundantly in /etc/portage/make.conf. Sometimes this is done out of ignorance, but it is also done to avoid flag filtering or flag replacing.
Flag filtering/replacing is done in many of the ebuilds in the Portage tree. It is usually done because packages fail to compile at certain -O
levels, or when the source code is too sensitive for any additional flags to be used. The ebuild will either filter out some or all CFLAGS and CXXFLAGS, or it may replace -O
with a different level.
The Gentoo Developer Manual outlines where and how flag filtering/replacing works.
It's possible to circumvent -O
filtering by redundantly listing the flags for a certain level, such as -O3
, by doing things like:
CFLAGS="-O3 -finline-functions -funswitch-loops"
However, this is not a smart thing to do. CFLAGS are filtered for a reason! When flags are filtered, it means that it is unsafe to build a package with those flags. Clearly, it is not safe to compile the whole system with -O3
if some of the flags turned on by that level will cause problems with certain packages. Therefore, don't try to "outsmart" the developers who maintain those packages. Trust the developers. Flag filtering and replacing is done to ensure stability of the system and application! If an ebuild specifies alternative flags, then don't try to get around it.
Building packages with unacceptable flags will most likely lead to problems. When reporting problems on Bugzilla, the flags that are used in /etc/portage/make.conf will be readily visible and developers will ask to recompile without those flags. Save the trouble of recompiling by not using redundant flags in the first place! Don't just automatically assume to be more knowledgeable than the developers.
What about LDFLAGS?
The Gentoo developers have already set basic, safe LDFLAGS in the base profiles, so they do not need to be changed.
Can I use per-package flags?
Using per-package flags complicates debugging and support. Make sure to mention the use of this feature in the bug reports together with the changes made.
Information on how to use per-package environment variables (including CFLAGS) is described in the Gentoo Handbook, "Per-Package Environment Variables".
Profile Guided Optimization (PGO)
Profile guided optimization (PGO) consists of compiling and profiling a program to assess hot paths in the code. Optimizations are then applied based on this analysis. Specifically, the code is compiled with -fprofile-generate
, which instrument the code. Second, the code is run with applications to collect profile information. Finally, using the profiled data, the code is compiled with -fprofile-use
. To manually enable PGO for packages, see this link.
Firefox also supports PGO although sometimes it may break the build.
Link Time Optimization (LTO)
LTO heavily increases compile times and if changing even one object file when compiling, LTO recompiles the whole code again. There is an ongoing GSoC project called "Bypass assembler when generating LTO object files" to make sure LTO only recompiles what it deems necessary.
LTO is still experimental. LTO may need to be disabled before reporting bugs because it is a common source of problems. The -flto
flag is used, with an optional auto
argument (Detects how many jobs to use) or an integer argument (An integer number of jobs to execute parallel).
See the LTO article for more information on LTO on Gentoo.
See also
- Configuring compile options (AMD64 Handbook)
- CPU_FLAGS_* — a USE_EXPAND variable containing instruction set and other CPU-specific features.
- Safe CFLAGS — a summary of "safe" settings for CFLAGS on Gentoo Linux.
- RUSTFLAGS
External resources
The following resources are of some help in further understanding optimization:
References
- ↑ GNU GCC Bugzilla, AVX/AVX2 no ymm registers used in a trivial reduction. Retrieved on 2017/07/18.
- ↑ GNU GCC Bugzilla, 'gcc -marc=native' sets L2 cache size equal to L3 cache size on i7 and i5 CPU. Retrieved on 2022/08/14.
- ↑ New 17.0 profiles in the Gentoo repository
- ↑ https://en.wikipedia.org/wiki/Full-employment_theorem
- ↑ https://gcc.gnu.org/install/build.html
This page is based on a document formerly found on our main website gentoo.org.
The following people contributed to the original document:
They are listed here because wiki history does not allow for any external attribution. If you edit the wiki article, please do not add yourself here; your contributions are recorded on each article's associated history page.