Clang Compiler User’s Manual

Introduction

The Clang Compiler is an open-source compiler for the C family of programming languages, aiming to be the best in class implementation of these languages. Clang builds on the LLVM optimizer and code generator, allowing it to provide high-quality optimization and code generation support for many targets. For more general information, please see the Clang Web Site or the LLVM Web Site.

This document describes important notes about using Clang as a compiler for an end-user, documenting the supported features, command line options, etc. If you are interested in using Clang to build a tool that processes code, please see “Clang” CFE Internals Manual. If you are interested in the Clang Static Analyzer, please see its web page.

Clang is one component in a complete toolchain for C family languages. A separate document describes the other pieces necessary to assemble a complete toolchain.

Clang is designed to support the C family of programming languages, which includes C, Objective-C, C++, and Objective-C++ as well as many dialects of those. For language-specific information, please see the corresponding language specific section:

In addition to these base languages and their dialects, Clang supports a broad variety of language extensions, which are documented in the corresponding language section. These extensions are provided to be compatible with the GCC, Microsoft, and other popular compilers as well as to improve functionality through Clang-specific features. The Clang driver and language features are intentionally designed to be as compatible with the GNU GCC compiler as reasonably possible, easing migration from GCC to Clang. In most cases, code “just works”. Clang also provides an alternative driver, clang-cl, that is designed to be compatible with the Visual C++ compiler, cl.exe.

In addition to language specific features, Clang has a variety of features that depend on what CPU architecture or operating system is being compiled for. Please see the Target-Specific Features and Limitations section for more details.

The rest of the introduction introduces some basic compiler terminology that is used throughout this manual and contains a basic introduction to using Clang as a command line compiler.

Terminology

Front end, parser, backend, preprocessor, undefined behavior, diagnostic, optimizer

Basic Usage

Intro to how to use a C compiler for newbies.

compile + link compile then link debug info enabling optimizations picking a language to use, defaults to C17 by default. Autosenses based on extension. using a makefile

Command Line Options

This section is generally an index into other sections. It does not go into depth on the ones that are covered by other sections. However, the first part introduces the language selection and other high level options like -c, -g, etc.

Options to Control Error and Warning Messages

-Werror

Turn warnings into errors.

-Werror=foo

Turn warning “foo” into an error.

-Wno-error=foo

Turn warning “foo” into a warning even if -Werror is specified.

-Wfoo

Enable warning “foo”. See the diagnostics reference for a complete list of the warning flags that can be specified in this way.

-Wno-foo

Disable warning “foo”.

-w

Disable all diagnostics.

-Weverything

Enable all diagnostics.

-pedantic

Warn on language extensions.

-pedantic-errors

Error on language extensions.

-Wsystem-headers

Enable warnings from system headers.

-ferror-limit=123

Stop emitting diagnostics after 123 errors have been produced. The default is 20, and the error limit can be disabled with -ferror-limit=0.

-ftemplate-backtrace-limit=123

Only emit up to 123 template instantiation notes within the template instantiation backtrace for a single warning or error. The default is 10, and the limit can be disabled with -ftemplate-backtrace-limit=0.

--warning-suppression-mappings=foo.txt

Suppress certain diagnostics for certain files.

Formatting of Diagnostics

Clang aims to produce beautiful diagnostics by default, particularly for new users that first come to Clang. However, different people have different preferences, and sometimes Clang is driven not by a human, but by a program that wants consistent and easily parsable output. For these cases, Clang provides a wide range of options to control the exact output format of the diagnostics that it generates.

-f[no-]show-column

Print column number in diagnostic.

This option, which defaults to on, controls whether or not Clang prints the column number of a diagnostic. For example, when this is enabled, Clang will print something like:

test.c:28:8: warning: extra tokens at end of #endif directive [-Wextra-tokens]
#endif bad
       ^
       //

When this is disabled, Clang will print “test.c:28: warning…” with no column number.

The printed column numbers count bytes from the beginning of the line; take care if your source contains multibyte characters.

-f[no-]show-source-location

Print source file/line/column information in diagnostic.

This option, which defaults to on, controls whether or not Clang prints the filename, line number and column number of a diagnostic. For example, when this is enabled, Clang will print something like:

test.c:28:8: warning: extra tokens at end of #endif directive [-Wextra-tokens]
#endif bad
       ^
       //

When this is disabled, Clang will not print the “test.c:28:8: ” part.

-f[no-]caret-diagnostics

Print source line and ranges from source code in diagnostic. This option, which defaults to on, controls whether or not Clang prints the source line, source ranges, and caret when emitting a diagnostic. For example, when this is enabled, Clang will print something like:

test.c:28:8: warning: extra tokens at end of #endif directive [-Wextra-tokens]
#endif bad
       ^
       //
-f[no-]color-diagnostics

This option, which defaults to on when a color-capable terminal is detected, controls whether or not Clang prints diagnostics in color.

When this option is enabled, Clang will use colors to highlight specific parts of the diagnostic, e.g.,

  test.c:28:8: warning: extra tokens at end of #endif directive [-Wextra-tokens]
  #endif bad
         ^
         //

When this is disabled, Clang will just print:

test.c:2:8: warning: extra tokens at end of #endif directive [-Wextra-tokens]
#endif bad
       ^
       //

If the NO_COLOR environment variable is defined and not empty (regardless of value), color diagnostics are disabled. If NO_COLOR is defined and -fcolor-diagnostics is passed on the command line, Clang will honor the command line argument.

-fansi-escape-codes

Controls whether ANSI escape codes are used instead of the Windows Console API to output colored diagnostics. This option is only used on Windows and defaults to off.

-fdiagnostics-format=clang/msvc/vi

Changes diagnostic output format to better match IDEs and command line tools.

This option controls the output format of the filename, line number, and column printed in diagnostic messages. The options, and their affect on formatting a simple conversion diagnostic, follow:

clang (default)
t.c:3:11: warning: conversion specifies type 'char *' but the argument has type 'int'
msvc
t.c(3,11) : warning: conversion specifies type 'char *' but the argument has type 'int'
vi
t.c +3:11: warning: conversion specifies type 'char *' but the argument has type 'int'
-f[no-]diagnostics-show-option

Enable [-Woption] information in diagnostic line.

This option, which defaults to on, controls whether or not Clang prints the associated warning group option name when outputting a warning diagnostic. For example, in this output:

test.c:28:8: warning: extra tokens at end of #endif directive [-Wextra-tokens]
#endif bad
       ^
       //

Passing -fno-diagnostics-show-option will prevent Clang from printing the [-Wextra-tokens] information in the diagnostic. This information tells you the flag needed to enable or disable the diagnostic, either from the command line or through #pragma GCC diagnostic.

-fdiagnostics-show-category=none/id/name

Enable printing category information in diagnostic line.

This option, which defaults to “none”, controls whether or not Clang prints the category associated with a diagnostic when emitting it. Each diagnostic may or many not have an associated category, if it has one, it is listed in the diagnostic categorization field of the diagnostic line (in the []’s).

For example, a format string warning will produce these three renditions based on the setting of this option:

t.c:3:11: warning: conversion specifies type 'char *' but the argument has type 'int' [-Wformat]
t.c:3:11: warning: conversion specifies type 'char *' but the argument has type 'int' [-Wformat,1]
t.c:3:11: warning: conversion specifies type 'char *' but the argument has type 'int' [-Wformat,Format String]

This category can be used by clients that want to group diagnostics by category, so it should be a high level category. We want dozens of these, not hundreds or thousands of them.

-f[no-]save-optimization-record[=<format>]

Enable optimization remarks during compilation and write them to a separate file.

This option, which defaults to off, controls whether Clang writes optimization reports to a separate file. By recording diagnostics in a file, users can parse or sort the remarks in a convenient way.

By default, the serialization format is YAML.

The supported serialization formats are:

  • -fsave-optimization-record=yaml: A structured YAML format.

  • -fsave-optimization-record=bitstream: A binary format based on LLVM Bitstream.

The output file is controlled by -foptimization-record-file.

In the absence of an explicit output file, the file is chosen using the following scheme:

<base>.opt.<format>

where <base> is based on the output file of the compilation (whether it’s explicitly specified through -o or not) when used with -c or -S. For example:

  • clang -fsave-optimization-record -c in.c -o out.o will generate out.opt.yaml

  • clang -fsave-optimization-record -c in.c will generate in.opt.yaml

When targeting (Thin)LTO, the base is derived from the output filename, and the extension is not dropped.

When targeting ThinLTO, the following scheme is used:

<base>.opt.<format>.thin.<num>.<format>

Darwin-only: when used for generating a linked binary from a source file (through an intermediate object file), the driver will invoke cc1 to generate a temporary object file. The temporary remark file will be emitted next to the object file, which will then be picked up by dsymutil and emitted in the .dSYM bundle. This is available for all formats except YAML.

For example:

clang -fsave-optimization-record=bitstream in.c -o out will generate

  • /var/folders/43/9y164hh52tv_2nrdxrj31nyw0000gn/T/a-9be59b.o

  • /var/folders/43/9y164hh52tv_2nrdxrj31nyw0000gn/T/a-9be59b.opt.bitstream

  • out

  • out.dSYM/Contents/Resources/Remarks/out

Darwin-only: compiling for multiple architectures will use the following scheme:

<base>-<arch>.opt.<format>

Note that this is incompatible with passing the -foptimization-record-file option.

-foptimization-record-file

Control the file to which optimization reports are written. This implies -fsave-optimization-record.

On Darwin platforms, this is incompatible with passing multiple -arch <arch> options.

-foptimization-record-passes

Only include passes which match a specified regular expression.

When optimization reports are being output (see -fsave-optimization-record), this option controls the passes that will be included in the final report.

If this option is not used, all the passes are included in the optimization record.

-f[no-]diagnostics-show-hotness

Enable profile hotness information in diagnostic line.

This option controls whether Clang prints the profile hotness associated with diagnostics in the presence of profile-guided optimization information. This is currently supported with optimization remarks (see Options to Emit Optimization Reports). The hotness information allows users to focus on the hot optimization remarks that are likely to be more relevant for run-time performance.

For example, in this output, the block containing the callsite of foo was executed 3000 times according to the profile data:

s.c:7:10: remark: foo inlined into bar (hotness: 3000) [-Rpass-analysis=inline]
  sum += foo(x, x - 2);
         ^

This option is implied when -fsave-optimization-record is used. Otherwise, it defaults to off.

-fdiagnostics-hotness-threshold

Prevent optimization remarks from being output if they do not have at least this hotness value.

This option, which defaults to zero, controls the minimum hotness an optimization remark would need in order to be output by Clang. This is currently supported with optimization remarks (see Options to Emit Optimization Reports) when profile hotness information in diagnostics is enabled (see -fdiagnostics-show-hotness).

-f[no-]diagnostics-fixit-info

Enable “FixIt” information in the diagnostics output.

This option, which defaults to on, controls whether or not Clang prints the information on how to fix a specific diagnostic underneath it when it knows. For example, in this output:

test.c:28:8: warning: extra tokens at end of #endif directive [-Wextra-tokens]
#endif bad
       ^
       //

Passing -fno-diagnostics-fixit-info will prevent Clang from printing the “//” line at the end of the message. This information is useful for users who may not understand what is wrong, but can be confusing for machine parsing.

-fdiagnostics-print-source-range-info

Print machine parsable information about source ranges. This option makes Clang print information about source ranges in a machine parsable format after the file/line/column number information. The information is a simple sequence of brace enclosed ranges, where each range lists the start and end line/column locations. For example, in this output:

exprs.c:47:15:{47:8-47:14}{47:17-47:24}: error: invalid operands to binary expression ('int *' and '_Complex float')
   P = (P-42) + Gamma*4;
       ~~~~~~ ^ ~~~~~~~

The {}’s are generated by -fdiagnostics-print-source-range-info.

The printed column numbers count bytes from the beginning of the line; take care if your source contains multibyte characters.

-fdiagnostics-parseable-fixits

Print Fix-Its in a machine parseable form.

This option makes Clang print available Fix-Its in a machine parseable format at the end of diagnostics. The following example illustrates the format:

fix-it:"t.cpp":{7:25-7:29}:"Gamma"

The range printed is a half-open range, so in this example the characters at column 25 up to but not including column 29 on line 7 in t.cpp should be replaced with the string “Gamma”. Either the range or the replacement string may be empty (representing strict insertions and strict erasures, respectively). Both the file name and the insertion string escape backslash (as “\\”), tabs (as “\t”), newlines (as “\n”), double quotes(as “\””) and non-printable characters (as octal “\xxx”).

The printed column numbers count bytes from the beginning of the line; take care if your source contains multibyte characters.

-fno-elide-type

Turns off elision in template type printing.

The default for template type printing is to elide as many template arguments as possible, removing those which are the same in both template types, leaving only the differences. Adding this flag will print all the template arguments. If supported by the terminal, highlighting will still appear on differing arguments.

Default:

t.cc:4:5: note: candidate function not viable: no known conversion from 'vector<map<[...], map<float, [...]>>>' to 'vector<map<[...], map<double, [...]>>>' for 1st argument;

-fno-elide-type:

t.cc:4:5: note: candidate function not viable: no known conversion from 'vector<map<int, map<float, int>>>' to 'vector<map<int, map<double, int>>>' for 1st argument;
-fdiagnostics-show-template-tree

Template type diffing prints a text tree.

For diffing large templated types, this option will cause Clang to display the templates as an indented text tree, one argument per line, with differences marked inline. This is compatible with -fno-elide-type.

Default:

t.cc:4:5: note: candidate function not viable: no known conversion from 'vector<map<[...], map<float, [...]>>>' to 'vector<map<[...], map<double, [...]>>>' for 1st argument;

With -fdiagnostics-show-template-tree:

t.cc:4:5: note: candidate function not viable: no known conversion for 1st argument;
  vector<
    map<
      [...],
      map<
        [float != double],
        [...]>>>
-fcaret-diagnostics-max-lines:

Controls how many lines of code clang prints for diagnostics. By default, clang prints a maximum of 16 lines of code.

-fdiagnostics-show-line-numbers:

Controls whether clang will print a margin containing the line number on the left of each line of code it prints for diagnostics.

Default:

test.cpp:5:1: error: 'main' must return 'int'
    5 | void main() {}
      | ^~~~
      | int

With -fno-diagnostics-show-line-numbers:

test.cpp:5:1: error: 'main' must return 'int'
void main() {}
^~~~
int

Individual Warning Groups

TODO: Generate this from tblgen. Define one anchor per warning group.

-Wextra-tokens

Warn about excess tokens at the end of a preprocessor directive.

This option, which defaults to on, enables warnings about extra tokens at the end of preprocessor directives. For example:

test.c:28:8: warning: extra tokens at end of #endif directive [-Wextra-tokens]
#endif bad
       ^

These extra tokens are not strictly conforming, and are usually best handled by commenting them out.

-Wambiguous-member-template

Warn about unqualified uses of a member template whose name resolves to another template at the location of the use.

This option, which defaults to on, enables a warning in the following code:

template<typename T> struct set{};
template<typename T> struct trait { typedef const T& type; };
struct Value {
  template<typename T> void set(typename trait<T>::type value) {}
};
void foo() {
  Value v;
  v.set<double>(3.2);
}

C++ [basic.lookup.classref] requires this to be an error, but, because it’s hard to work around, Clang downgrades it to a warning as an extension.

-Wbind-to-temporary-copy

Warn about an unusable copy constructor when binding a reference to a temporary.

This option enables warnings about binding a reference to a temporary when the temporary doesn’t have a usable copy constructor. For example:

struct NonCopyable {
  NonCopyable();
private:
  NonCopyable(const NonCopyable&);
};
void foo(const NonCopyable&);
void bar() {
  foo(NonCopyable());  // Disallowed in C++98; allowed in C++11.
}

struct NonCopyable2 {
  NonCopyable2();
  NonCopyable2(NonCopyable2&);
};
void foo(const NonCopyable2&);
void bar() {
  foo(NonCopyable2());  // Disallowed in C++98; allowed in C++11.
}

Note that if NonCopyable2::NonCopyable2() has a default argument whose instantiation produces a compile error, that error will still be a hard error in C++98 mode even if this warning is turned off.

Options to Control Clang Crash Diagnostics

As unbelievable as it may sound, Clang does crash from time to time. Generally, this only occurs to those living on the bleeding edge. Clang goes to great lengths to assist you in filing a bug report. Specifically, Clang generates preprocessed source file(s) and associated run script(s) upon a crash. These files should be attached to a bug report to ease reproducibility of the failure. Below are the command line options to control the crash diagnostics.

-fcrash-diagnostics=<val>

Valid values are:

  • off (Disable auto-generation of preprocessed source files during a clang crash.)

  • compiler (Generate diagnostics for compiler crashes (default))

  • all (Generate diagnostics for all tools which support it)

-fno-crash-diagnostics

Disable auto-generation of preprocessed source files during a clang crash.

The -fno-crash-diagnostics flag can be helpful for speeding the process of generating a delta reduced test case.

-fcrash-diagnostics-dir=<dir>

Specify where to write the crash diagnostics files; defaults to the usual location for temporary files.

CLANG_CRASH_DIAGNOSTICS_DIR=<dir>

Like -fcrash-diagnostics-dir=<dir>, specifies where to write the crash diagnostics files, but with lower precedence than the option.

Clang is also capable of generating preprocessed source file(s) and associated run script(s) even without a crash. This is specially useful when trying to generate a reproducer for warnings or errors while using modules.

-gen-reproducer

Generates preprocessed source files, a reproducer script and if relevant, a cache containing: built module pcm’s and all headers needed to rebuild the same modules.

Options to Emit Optimization Reports

Optimization reports trace, at a high-level, all the major decisions done by compiler transformations. For instance, when the inliner decides to inline function foo() into bar(), or the loop unroller decides to unroll a loop N times, or the vectorizer decides to vectorize a loop body.

Clang offers a family of flags which the optimizers can use to emit a diagnostic in three cases:

  1. When the pass makes a transformation (-Rpass).

  2. When the pass fails to make a transformation (-Rpass-missed).

  3. When the pass determines whether or not to make a transformation (-Rpass-analysis).

NOTE: Although the discussion below focuses on -Rpass, the exact same options apply to -Rpass-missed and -Rpass-analysis.

Since there are dozens of passes inside the compiler, each of these flags take a regular expression that identifies the name of the pass which should emit the associated diagnostic. For example, to get a report from the inliner, compile the code with:

$ clang -O2 -Rpass=inline code.cc -o code
code.cc:4:25: remark: foo inlined into bar [-Rpass=inline]
int bar(int j) { return foo(j, j - 2); }
                        ^

Note that remarks from the inliner are identified with [-Rpass=inline]. To request a report from every optimization pass, you should use -Rpass=.* (in fact, you can use any valid POSIX regular expression). However, do not expect a report from every transformation made by the compiler. Optimization remarks do not really make sense outside of the major transformations (e.g., inlining, vectorization, loop optimizations) and not every optimization pass supports this feature.

Note that when using profile-guided optimization information, profile hotness information can be included in the remarks (see -fdiagnostics-show-hotness).

Current limitations

  1. Optimization remarks that refer to function names will display the mangled name of the function. Since these remarks are emitted by the back end of the compiler, it does not know anything about the input language, nor its mangling rules.

  2. Some source locations are not displayed correctly. The front end has a more detailed source location tracking than the locations included in the debug info (e.g., the front end can locate code inside macro expansions). However, the locations used by -Rpass are translated from debug annotations. That translation can be lossy, which results in some remarks having no location information.

Options to Emit Resource Consumption Reports

These are options that report execution time and consumed memory of different compilations steps.

-fproc-stat-report=

This option requests driver to print used memory and execution time of each compilation step. The clang driver during execution calls different tools, like compiler, assembler, linker etc. With this option the driver reports total execution time, the execution time spent in user mode and peak memory usage of each the called tool. Value of the option specifies where the report is sent to. If it specifies a regular file, the data are saved to this file in CSV format:

$ clang -fproc-stat-report=abc foo.c
$ cat abc
clang-11,"/tmp/foo-123456.o",92000,84000,87536
ld,"a.out",900,8000,53568

The data on each row represent:

  • file name of the tool executable,

  • output file name in quotes,

  • total execution time in microseconds,

  • execution time in user mode in microseconds,

  • peak memory usage in Kb.

It is possible to specify this option without any value. In this case statistics are printed on standard output in human readable format:

$ clang -fproc-stat-report foo.c
clang-11: output=/tmp/foo-855a8e.o, total=68.000 ms, user=60.000 ms, mem=86920 Kb
ld: output=a.out, total=8.000 ms, user=4.000 ms, mem=52320 Kb

The report file specified in the option is locked for write, so this option can be used to collect statistics in parallel builds. The report file is not cleared, new data is appended to it, thus making possible to accumulate build statistics.

You can also use environment variables to control the process statistics reporting. Setting CC_PRINT_PROC_STAT to 1 enables the feature, the report goes to stdout in human readable format. Setting CC_PRINT_PROC_STAT_FILE to a fully qualified file path makes it report process statistics to the given file in the CSV format. Specifying a relative path will likely lead to multiple files with the same name created in different directories, since the path is relative to a changing working directory.

These environment variables are handy when you need to request the statistics report without changing your build scripts or alter the existing set of compiler options. Note that -fproc-stat-report take precedence over CC_PRINT_PROC_STAT and CC_PRINT_PROC_STAT_FILE.

$ export CC_PRINT_PROC_STAT=1
$ export CC_PRINT_PROC_STAT_FILE=~/project-build-proc-stat.csv
$ make

Other Options

Clang options that don’t fit neatly into other categories.

-fgnuc-version=

This flag controls the value of __GNUC__ and related macros. This flag does not enable or disable any GCC extensions implemented in Clang. Setting the version to zero causes Clang to leave __GNUC__ and other GNU-namespaced macros, such as __GXX_WEAK__, undefined.

-MV

When emitting a dependency file, use formatting conventions appropriate for NMake or Jom. Ignored unless another option causes Clang to emit a dependency file.

When Clang emits a dependency file (e.g., you supplied the -M option) most filenames can be written to the file without any special formatting. Different Make tools will treat different sets of characters as “special” and use different conventions for telling the Make tool that the character is actually part of the filename. Normally Clang uses backslash to “escape” a special character, which is the convention used by GNU Make. The -MV option tells Clang to put double-quotes around the entire filename, which is the convention used by NMake and Jom.

-femit-dwarf-unwind=<value>

When to emit DWARF unwind (EH frame) info. This is a Mach-O-specific option.

Valid values are:

  • no-compact-unwind - Only emit DWARF unwind when compact unwind encodings aren’t available. This is the default for arm64.

  • always - Always emit DWARF unwind regardless.

  • default - Use the platform-specific default (always for all non-arm64-platforms).

no-compact-unwind is a performance optimization – Clang will emit smaller object files that are more quickly processed by the linker. This may cause binary compatibility issues on older x86_64 targets, however, so use it with caution.

-fdisable-block-signature-string

Instruct clang not to emit the signature string for blocks. Disabling the string can potentially break existing code that relies on it. Users should carefully consider this possibiilty when using the flag.

Configuration files

Configuration files group command-line options and allow all of them to be specified just by referencing the configuration file. They may be used, for example, to collect options required to tune compilation for particular target, such as -L, -I, -l, --sysroot, codegen options, etc.

Configuration files can be either specified on the command line or loaded from default locations. If both variants are present, the default configuration files are loaded first.

The command line option --config= can be used to specify explicit configuration files in a Clang invocation. If the option is used multiple times, all specified files are loaded, in order. For example:

clang --config=/home/user/cfgs/testing.txt
clang --config=debug.cfg --config=runtimes.cfg

If the provided argument contains a directory separator, it is considered as a file path, and options are read from that file. Otherwise the argument is treated as a file name and is searched for sequentially in the directories:

  • user directory,

  • system directory,

  • the directory where Clang executable resides.

Both user and system directories for configuration files can be specified either during build or during runtime. At build time, use CLANG_CONFIG_FILE_USER_DIR and CLANG_CONFIG_FILE_SYSTEM_DIR. At run time use the --config-user-dir= and --config-system-dir= command line options. Specifying config directories at runtime overrides the config directories set at build time The first file found is used. It is an error if the required file cannot be found.

The default configuration files are searched for in the same directories following the rules described in the next paragraphs. Loading default configuration files can be disabled entirely via passing the --no-default-config flag.

First, the algorithm searches for a configuration file named <triple>-<driver>.cfg where triple is the triple for the target being built for, and driver is the name of the currently used driver. The algorithm first attempts to use the canonical name for the driver used, then falls back to the one found in the executable name.

The following canonical driver names are used:

  • clang for the gcc driver (used to compile C programs)

  • clang++ for the gxx driver (used to compile C++ programs)

  • clang-cpp for the cpp driver (pure preprocessor)

  • clang-cl for the cl driver

  • flang for the flang driver

  • clang-dxc for the dxc driver

For example, when calling x86_64-pc-linux-gnu-clang-g++, the driver will first attempt to use the configuration file named:

x86_64-pc-linux-gnu-clang++.cfg

If this file is not found, it will attempt to use the name found in the executable instead:

x86_64-pc-linux-gnu-clang-g++.cfg

Note that options such as --driver-mode=, --target=, -m32 affect the search algorithm. For example, the aforementioned executable called with -m32 argument will instead search for:

i386-pc-linux-gnu-clang++.cfg

If none of the aforementioned files are found, the driver will instead search for separate driver and target configuration files and attempt to load both. The former is named <driver>.cfg while the latter is named <triple>.cfg. Similarly to the previous variants, the canonical driver name will be preferred, and the compiler will fall back to the actual name.

For example, x86_64-pc-linux-gnu-clang-g++ will attempt to load two configuration files named respectively:

clang++.cfg
x86_64-pc-linux-gnu.cfg

with fallback to trying:

clang-g++.cfg
x86_64-pc-linux-gnu.cfg

It is not an error if either of these files is not found.

The configuration file consists of command-line options specified on one or more lines. Lines composed of whitespace characters only are ignored as well as lines in which the first non-blank character is #. Long options may be split between several lines by a trailing backslash. Here is example of a configuration file:

# Several options on line
-c --target=x86_64-unknown-linux-gnu

# Long option split between lines
-I/usr/lib/gcc/x86_64-linux-gnu/5.4.0/../../../../\
include/c++/5.4.0

# other config files may be included
@linux.options

Files included by @file directives in configuration files are resolved relative to the including file. For example, if a configuration file ~/.llvm/target.cfg contains the directive @os/linux.opts, the file linux.opts is searched for in the directory ~/.llvm/os. Another way to include a file content is using the command line option --config=. It works similarly but the included file is searched for using the rules for configuration files.

To generate paths relative to the configuration file, the <CFGDIR> token may be used. This will expand to the absolute path of the directory containing the configuration file.

In cases where a configuration file is deployed alongside SDK contents, the SDK directory can remain fully portable by using <CFGDIR> prefixed paths. In this way, the user may only need to specify a root configuration file with --config= to establish every aspect of the SDK with the compiler:

--target=foo
-isystem <CFGDIR>/include
-L <CFGDIR>/lib
-T <CFGDIR>/ldscripts/link.ld

Language and Target-Independent Features

Controlling Errors and Warnings

Clang provides a number of ways to control which code constructs cause it to emit errors and warning messages, and how they are displayed to the console.

Controlling How Clang Displays Diagnostics

When Clang emits a diagnostic, it includes rich information in the output, and gives you fine-grain control over which information is printed. Clang has the ability to print this information, and these are the options that control it:

  1. A file/line/column indicator that shows exactly where the diagnostic occurs in your code [-fshow-column, -fshow-source-location].

  2. A categorization of the diagnostic as a note, warning, error, or fatal error.

  3. A text string that describes what the problem is.

  4. An option that indicates how to control the diagnostic (for diagnostics that support it) [-fdiagnostics-show-option].

  5. A high-level category for the diagnostic for clients that want to group diagnostics by class (for diagnostics that support it) [-fdiagnostics-show-category].

  6. The line of source code that the issue occurs on, along with a caret and ranges that indicate the important locations [-fcaret-diagnostics].

  7. “FixIt” information, which is a concise explanation of how to fix the problem (when Clang is certain it knows) [-fdiagnostics-fixit-info].

  8. A machine-parsable representation of the ranges involved (off by default) [-fdiagnostics-print-source-range-info].

For more information please see Formatting of Diagnostics.

Diagnostic Mappings

All diagnostics are mapped into one of these 6 classes:

  • Ignored

  • Note

  • Remark

  • Warning

  • Error

  • Fatal

Diagnostic Categories

Though not shown by default, diagnostics may each be associated with a high-level category. This category is intended to make it possible to triage builds that produce a large number of errors or warnings in a grouped way.

Categories are not shown by default, but they can be turned on with the -fdiagnostics-show-category option. When set to “name”, the category is printed textually in the diagnostic output. When it is set to “id”, a category number is printed. The mapping of category names to category id’s can be obtained by running ‘clang   --print-diagnostic-categories’.

Controlling Diagnostics via Command Line Flags

TODO: -W flags, -pedantic, etc

Controlling Diagnostics via Pragmas

Clang can also control what diagnostics are enabled through the use of pragmas in the source code. This is useful for turning off specific warnings in a section of source code. Clang supports GCC’s pragma for compatibility with existing source code, so #pragma GCC diagnostic and #pragma clang diagnostic are synonyms for Clang. GCC will ignore #pragma clang diagnostic, though.

The pragma may control any warning that can be used from the command line. Warnings may be set to ignored, warning, error, or fatal. The following example code will tell Clang or GCC to ignore the -Wall warnings:

#pragma GCC diagnostic ignored "-Wall"

Clang also allows you to push and pop the current warning state. This is particularly useful when writing a header file that will be compiled by other people, because you don’t know what warning flags they build with.

In the below example -Wextra-tokens is ignored for only a single line of code, after which the diagnostics return to whatever state had previously existed.

#if foo
#endif foo // warning: extra tokens at end of #endif directive

#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wextra-tokens"

#if foo
#endif foo // no warning

#pragma GCC diagnostic pop

The push and pop pragmas will save and restore the full diagnostic state of the compiler, regardless of how it was set. It should be noted that while Clang supports the GCC pragma, Clang and GCC do not support the exact same set of warnings, so even when using GCC compatible #pragmas there is no guarantee that they will have identical behaviour on both compilers.

Clang also doesn’t yet support GCC behavior for #pragma diagnostic pop that doesn’t have a corresponding #pragma diagnostic push. In this case GCC pretends that there is a #pragma diagnostic push at the very beginning of the source file, so “unpaired” #pragma diagnostic pop matches that implicit push. This makes a difference for #pragma GCC diagnostic ignored which are not guarded by push and pop. Refer to GCC documentation for details.

Like GCC, Clang accepts ignored, warning, error, and fatal severity levels. They can be used to change severity of a particular diagnostic for a region of source file. A notable difference from GCC is that diagnostic not enabled via command line arguments can’t be enabled this way yet.

Some diagnostics associated with a -W flag have the error severity by default. They can be ignored or downgraded to warnings:

// C only
#pragma GCC diagnostic warning "-Wimplicit-function-declaration"
int main(void) { puts(""); }

In addition to controlling warnings and errors generated by the compiler, it is possible to generate custom warning and error messages through the following pragmas:

// The following will produce warning messages
#pragma message "some diagnostic message"
#pragma GCC warning "TODO: replace deprecated feature"

// The following will produce an error message
#pragma GCC error "Not supported"

These pragmas operate similarly to the #warning and #error preprocessor directives, except that they may also be embedded into preprocessor macros via the C99 _Pragma operator, for example:

#define STR(X) #X
#define DEFER(M,...) M(__VA_ARGS__)
#define CUSTOM_ERROR(X) _Pragma(STR(GCC error(X " at line " DEFER(STR,__LINE__))))

CUSTOM_ERROR("Feature not available");

Controlling Diagnostics in System Headers

Warnings are suppressed when they occur in system headers. By default, an included file is treated as a system header if it is found in an include path specified by -isystem, but this can be overridden in several ways.

The system_header pragma can be used to mark the current file as being a system header. No warnings will be produced from the location of the pragma onwards within the same file.

#if foo
#endif foo // warning: extra tokens at end of #endif directive

#pragma clang system_header

#if foo
#endif foo // no warning

The –system-header-prefix= and –no-system-header-prefix= command-line arguments can be used to override whether subsets of an include path are treated as system headers. When the name in a #include directive is found within a header search path and starts with a system prefix, the header is treated as a system header. The last prefix on the command-line which matches the specified header name takes precedence. For instance:

$ clang -Ifoo -isystem bar --system-header-prefix=x/ \
    --no-system-header-prefix=x/y/

Here, #include "x/a.h" is treated as including a system header, even if the header is found in foo, and #include "x/y/b.h" is treated as not including a system header, even if the header is found in bar.

A #include directive which finds a file relative to the current directory is treated as including a system header if the including file is treated as a system header.

Controlling Deprecation Diagnostics in Clang-Provided C Runtime Headers

Clang is responsible for providing some of the C runtime headers that cannot be provided by a platform CRT, such as implementation limits or when compiling in freestanding mode. Define the _CLANG_DISABLE_CRT_DEPRECATION_WARNINGS macro prior to including such a C runtime header to disable the deprecation warnings. Note that the C Standard Library headers are allowed to transitively include other standard library headers (see 7.1.2p5), and so the most appropriate use of this macro is to set it within the build system using -D or before any include directives in the translation unit.

#define _CLANG_DISABLE_CRT_DEPRECATION_WARNINGS
#include <stdint.h>    // Clang CRT deprecation warnings are disabled.
#include <stdatomic.h> // Clang CRT deprecation warnings are disabled.

Enabling All Diagnostics

In addition to the traditional -W flags, one can enable all diagnostics by passing -Weverything. This works as expected with -Werror, and also includes the warnings from -pedantic. Some diagnostics contradict each other, therefore, users of -Weverything often disable many diagnostics such as -Wno-c++98-compat and -Wno-c++-compat because they contradict recent C++ standards.

Since -Weverything enables every diagnostic, we generally don’t recommend using it. -Wall -Wextra are a better choice for most projects. Using -Weverything means that updating your compiler is more difficult because you’re exposed to experimental diagnostics which might be of lower quality than the default ones. If you do use -Weverything then we advise that you address all new compiler diagnostics as they get added to Clang, either by fixing everything they find or explicitly disabling that diagnostic with its corresponding Wno- option.

Note that when combined with -w (which disables all warnings), disabling all warnings wins.

Controlling Diagnostics via Suppression Mappings

Warning suppression mappings enable users to suppress Clang’s diagnostics at a per-file granularity. This allows enforcing diagnostics in specific parts of the project even if there are violations in some headers.

$ cat mappings.txt
[unused]
src:foo/*

$ clang --warning-suppression-mappings=mapping.txt -Wunused foo/bar.cc
# This compilation won't emit any unused findings for sources under foo/
# directory. But it'll still complain for all the other sources, e.g:
$ cat foo/bar.cc
#include "dir/include.h" // Clang flags unused declarations here.
#include "foo/include.h" // but unused warnings under this source is omitted.
#include "next_to_bar_cc.h" // as are unused warnings from this header file.
// Further, unused warnings in the remainder of bar.cc are also omitted.

See Warning suppression mappings for details about the file format and functionality.

Controlling Static Analyzer Diagnostics

While not strictly part of the compiler, the diagnostics from Clang’s static analyzer can also be influenced by the user via changes to the source code. See the available annotations and the analyzer’s FAQ page for more information.

Precompiled Headers

Precompiled headers are a general approach employed by many compilers to reduce compilation time. The underlying motivation of the approach is that it is common for the same (and often large) header files to be included by multiple source files. Consequently, compile times can often be greatly improved by caching some of the (redundant) work done by a compiler to process headers. Precompiled header files, which represent one of many ways to implement this optimization, are literally files that represent an on-disk cache that contains the vital information necessary to reduce some of the work needed to process a corresponding header file. While details of precompiled headers vary between compilers, precompiled headers have been shown to be highly effective at speeding up program compilation on systems with very large system headers (e.g., macOS).

Generating a PCH File

To generate a PCH file using Clang, one invokes Clang with the -x <language>-header option. This mirrors the interface in GCC for generating PCH files:

$ gcc -x c-header test.h -o test.h.gch
$ clang -x c-header test.h -o test.h.pch

Using a PCH File

A PCH file can then be used as a prefix header when a -include-pch option is passed to clang:

$ clang -include-pch test.h.pch test.c -o test

The clang driver will check if the PCH file test.h.pch is available; if so, the contents of test.h (and the files it includes) will be processed from the PCH file. Otherwise, Clang will report an error.

Note

Clang does not automatically use PCH files for headers that are directly included within a source file or indirectly via -include. For example:

$ clang -x c-header test.h -o test.h.pch
$ cat test.c
#include "test.h"
$ clang test.c -o test

In this example, clang will not automatically use the PCH file for test.h since test.h was included directly in the source file and not specified on the command line using -include-pch.

Relocatable PCH Files

It is sometimes necessary to build a precompiled header from headers that are not yet in their final, installed locations. For example, one might build a precompiled header within the build tree that is then meant to be installed alongside the headers. Clang permits the creation of “relocatable” precompiled headers, which are built with a given path (into the build directory) and can later be used from an installed location.

To build a relocatable precompiled header, place your headers into a subdirectory whose structure mimics the installed location. For example, if you want to build a precompiled header for the header mylib.h that will be installed into /usr/include, create a subdirectory build/usr/include and place the header mylib.h into that subdirectory. If mylib.h depends on other headers, then they can be stored within build/usr/include in a way that mimics the installed location.

Building a relocatable precompiled header requires two additional arguments. First, pass the --relocatable-pch flag to indicate that the resulting PCH file should be relocatable. Second, pass -isysroot /path/to/build, which makes all includes for your library relative to the build directory. For example:

# clang -x c-header --relocatable-pch -isysroot /path/to/build /path/to/build/mylib.h mylib.h.pch

When loading the relocatable PCH file, the various headers used in the PCH file are found from the system header root. For example, mylib.h can be found in /usr/include/mylib.h. If the headers are installed in some other system root, the -isysroot option can be used provide a different system root from which the headers will be based. For example, -isysroot /Developer/SDKs/MacOSX10.4u.sdk will look for mylib.h in /Developer/SDKs/MacOSX10.4u.sdk/usr/include/mylib.h.

Relocatable precompiled headers are intended to be used in a limited number of cases where the compilation environment is tightly controlled and the precompiled header cannot be generated after headers have been installed.

Controlling Floating Point Behavior

Clang provides a number of ways to control floating point behavior, including with command line options and source pragmas. This section describes the various floating point semantic modes and the corresponding options.

Floating Point Semantic Modes

Mode

Values

ffp-exception-behavior

{ignore, strict, maytrap}

fenv_access

{off, on}

(none)

frounding-math

{dynamic, tonearest, downward, upward, towardzero}

ffp-contract

{on, off, fast, fast-honor-pragmas}

fdenormal-fp-math

{IEEE, PreserveSign, PositiveZero}

fdenormal-fp-math-fp32

{IEEE, PreserveSign, PositiveZero}

fmath-errno

{on, off}

fhonor-nans

{on, off}

fhonor-infinities

{on, off}

fsigned-zeros

{on, off}

freciprocal-math

{on, off}

fallow-approximate-fns

{on, off}

fassociative-math

{on, off}

fcomplex-arithmetic

{basic, improved, full, promoted}

This table describes the option settings that correspond to the three floating point semantic models: precise (the default), strict, and fast.

Floating Point Models

Mode

Precise

Strict

Fast

Aggressive

except_behavior

ignore

strict

ignore

ignore

fenv_access

off

on

off

off

rounding_mode

tonearest

dynamic

tonearest

tonearest

contract

on

off

fast

fast

support_math_errno

on

on

off

off

no_honor_nans

off

off

off

on

no_honor_infinities

off

off

off

on

no_signed_zeros

off

off

on

on

allow_reciprocal

off

off

on

on

allow_approximate_fns

off

off

on

on

allow_reassociation

off

off

on

on

complex_arithmetic

full

full

promoted

basic

The -ffp-model option does not modify the fdenormal-fp-math setting, but it does have an impact on whether crtfastmath.o is linked. Because linking crtfastmath.o has a global effect on the program, and because the global denormal handling can be changed in other ways, the state of fdenormal-fp-math handling cannot be assumed in any function based on fp-model. See A note about crtfastmath.o for more details.

-ffast-math

Enable fast-math mode. This option lets the compiler make aggressive, potentially-lossy assumptions about floating-point math. These include:

  • Floating-point math obeys regular algebraic rules for real numbers (e.g. + and * are associative, x/y == x * (1/y), and (a + b) * c == a * c + b * c),

  • No NaN or infinite values will be operands or results of floating-point operations,

  • +0 and -0 may be treated as interchangeable.

-ffast-math also defines the __FAST_MATH__ preprocessor macro. Some math libraries recognize this macro and change their behavior. With the exception of -ffp-contract=fast, using any of the options below to disable any of the individual optimizations in -ffast-math will cause __FAST_MATH__ to no longer be set. -ffast-math enables -fcx-limited-range.

This option implies:

  • -fno-honor-infinities

  • -fno-honor-nans

  • -fapprox-func

  • -fno-math-errno

  • -ffinite-math-only

  • -fassociative-math

  • -freciprocal-math

  • -fno-signed-zeros

  • -fno-trapping-math

  • -fno-rounding-math

  • -ffp-contract=fast

Note: -ffast-math causes crtfastmath.o to be linked with code unless -shared or -mno-daz-ftz is present. See A note about crtfastmath.o for more details.

-fno-fast-math

Disable fast-math mode. This options disables unsafe floating-point optimizations by preventing the compiler from making any transformations that could affect the results.

This option implies:

  • -fhonor-infinities

  • -fhonor-nans

  • -fno-approx-func

  • -fno-finite-math-only

  • -fno-associative-math

  • -fno-reciprocal-math

  • -fsigned-zeros

  • -ffp-contract=on

Also, this option resets following options to their target-dependent defaults.

  • -f[no-]math-errno

There is ambiguity about how -ffp-contract, -ffast-math, and -fno-fast-math behave when combined. To keep the value of -ffp-contract consistent, we define this set of rules:

  • -ffast-math sets ffp-contract to fast.

  • -fno-fast-math sets -ffp-contract to on (fast for CUDA and HIP).

  • If -ffast-math and -ffp-contract are both seen, but -ffast-math is not followed by -fno-fast-math, ffp-contract will be given the value of whichever option was last seen.

  • If -fno-fast-math is seen and -ffp-contract has been seen at least once, the ffp-contract will get the value of the last seen value of -ffp-contract.

  • If -fno-fast-math is seen and -ffp-contract has not been seen, the -ffp-contract setting is determined by the default value of -ffp-contract.

Note: -fno-fast-math causes crtfastmath.o to not be linked with code unless -mdaz-ftz is present.

-fdenormal-fp-math=<value>

Select which denormal numbers the code is permitted to require.

Valid values are:

  • ieee - IEEE 754 denormal numbers

  • preserve-sign - the sign of a flushed-to-zero number is preserved in the sign of 0

  • positive-zero - denormals are flushed to positive zero

The default value depends on the target. For most targets, defaults to ieee.

-f[no-]strict-float-cast-overflow

When a floating-point value is not representable in a destination integer type, the code has undefined behavior according to the language standard. By default, Clang will not guarantee any particular result in that case. With the ‘no-strict’ option, Clang will saturate towards the smallest and largest representable integer values instead. NaNs will be converted to zero. Defaults to -fstrict-float-cast-overflow.

-f[no-]math-errno

Require math functions to indicate errors by setting errno. The default varies by ToolChain. -fno-math-errno allows optimizations that might cause standard C math functions to not set errno. For example, on some systems, the math function sqrt is specified as setting errno to EDOM when the input is negative. On these systems, the compiler cannot normally optimize a call to sqrt to use inline code (e.g. the x86 sqrtsd instruction) without additional checking to ensure that errno is set appropriately. -fno-math-errno permits these transformations.

On some targets, math library functions never set errno, and so -fno-math-errno is the default. This includes most BSD-derived systems, including Darwin.

-f[no-]trapping-math

Control floating point exception behavior. -fno-trapping-math allows optimizations that assume that floating point operations cannot generate traps such as divide-by-zero, overflow and underflow.

  • The option -ftrapping-math behaves identically to -ffp-exception-behavior=strict.

  • The option -fno-trapping-math behaves identically to -ffp-exception-behavior=ignore. This is the default.

-ffp-contract=<value>

Specify when the compiler is permitted to form fused floating-point operations, such as fused multiply-add (FMA). Fused operations are permitted to produce more precise results than performing the same operations separately.

The C standard permits intermediate floating-point results within an expression to be computed with more precision than their type would normally allow. This permits operation fusing, and Clang takes advantage of this by default. This behavior can be controlled with the FP_CONTRACT and clang fp contract pragmas. Please refer to the pragma documentation for a description of how the pragmas interact with this option.

Valid values are:

  • fast (fuse across statements disregarding pragmas, default for CUDA)

  • on (fuse in the same statement unless dictated by pragmas, default for languages other than CUDA/HIP)

  • off (never fuse)

  • fast-honor-pragmas (fuse across statements unless dictated by pragmas, default for HIP)

-f[no-]honor-infinities

Allow floating-point optimizations that assume arguments and results are not +-Inf. Defaults to -fhonor-infinities.

If both -fno-honor-infinities and -fno-honor-nans are used, has the same effect as specifying -ffinite-math-only.

-f[no-]honor-nans

Allow floating-point optimizations that assume arguments and results are not NaNs. Defaults to -fhonor-nans.

If both -fno-honor-infinities and -fno-honor-nans are used, has the same effect as specifying -ffinite-math-only.

-f[no-]approx-func

Allow certain math function calls (such as log, sqrt, pow, etc) to be replaced with an approximately equivalent set of instructions or alternative math function calls. For example, a pow(x, 0.25) may be replaced with sqrt(sqrt(x)), despite being an inexact result in cases where x is -0.0 or -inf. Defaults to -fno-approx-func.

-f[no-]signed-zeros

Allow optimizations that ignore the sign of floating point zeros. Defaults to -fsigned-zeros.

-f[no-]associative-math

Allow floating point operations to be reassociated. Defaults to -fno-associative-math.

-f[no-]reciprocal-math

Allow division operations to be transformed into multiplication by a reciprocal. This can be significantly faster than an ordinary division but can also have significantly less precision. Defaults to -fno-reciprocal-math.

-f[no-]unsafe-math-optimizations

Allow unsafe floating-point optimizations. -funsafe-math-optimizations also implies:

  • -fapprox-func

  • -fassociative-math

  • -freciprocal-math

  • -fno-signed-zeros

  • -fno-trapping-math

  • -ffp-contract=fast

-fno-unsafe-math-optimizations implies:

  • -fno-approx-func

  • -fno-associative-math

  • -fno-reciprocal-math

  • -fsigned-zeros

  • -ffp-contract=on

There is ambiguity about how -ffp-contract, -funsafe-math-optimizations, and -fno-unsafe-math-optimizations behave when combined. Explanation in -fno-fast-math also applies to these options.

Defaults to -fno-unsafe-math-optimizations.

-f[no-]finite-math-only

Allow floating-point optimizations that assume arguments and results are not NaNs or +-Inf. -ffinite-math-only defines the __FINITE_MATH_ONLY__ preprocessor macro. -ffinite-math-only implies:

  • -fno-honor-infinities

  • -fno-honor-nans

-ffno-inite-math-only implies:

  • -fhonor-infinities

  • -fhonor-nans

Defaults to -fno-finite-math-only.

-f[no-]rounding-math

Force floating-point operations to honor the dynamically-set rounding mode by default.

The result of a floating-point operation often cannot be exactly represented in the result type and therefore must be rounded. IEEE 754 describes different rounding modes that control how to perform this rounding, not all of which are supported by all implementations. C provides interfaces (fesetround and fesetenv) for dynamically controlling the rounding mode, and while it also recommends certain conventions for changing the rounding mode, these conventions are not typically enforced in the ABI. Since the rounding mode changes the numerical result of operations, the compiler must understand something about it in order to optimize floating point operations.

Note that floating-point operations performed as part of constant initialization are formally performed prior to the start of the program and are therefore not subject to the current rounding mode. This includes the initialization of global variables and local static variables. Floating-point operations in these contexts will be rounded using FE_TONEAREST.

  • The option -fno-rounding-math allows the compiler to assume that the rounding mode is set to FE_TONEAREST. This is the default.

  • The option -frounding-math forces the compiler to honor the dynamically-set rounding mode. This prevents optimizations which might affect results if the rounding mode changes or is different from the default; for example, it prevents floating-point operations from being reordered across most calls and prevents constant-folding when the result is not exactly representable.

-ffp-model=<value>

Specify floating point behavior. -ffp-model is an umbrella option that encompasses functionality provided by other, single purpose, floating point options. Valid values are: precise, strict, fast, and aggressive. Details:

  • precise Disables optimizations that are not value-safe on floating-point data, although FP contraction (FMA) is enabled (-ffp-contract=on). This is the default behavior. This value resets -fmath-errno to its target-dependent default.

  • strict Enables -frounding-math and -ffp-exception-behavior=strict, and disables contractions (FMA). All of the -ffast-math enablements are disabled. Enables STDC FENV_ACCESS: by default FENV_ACCESS is disabled. This option setting behaves as though #pragma STDC FENV_ACCESS ON appeared at the top of the source file.

  • fast Behaves identically to specifying -funsafe-math-optimizations, -fno-math-errno and -fcomplex-arithmetic=promoted ffp-contract=fast

  • aggressive Behaves identically to specifying both -ffast-math and ffp-contract=fast

Note: If your command line specifies multiple instances of the -ffp-model option, or if your command line option specifies -ffp-model and later on the command line selects a floating point option that has the effect of negating part of the ffp-model that has been selected, then the compiler will issue a diagnostic warning that the override has occurred.

-ffp-exception-behavior=<value>

Specify the floating-point exception behavior.

Valid values are: ignore, maytrap, and strict. The default value is ignore. Details:

  • ignore The compiler assumes that the exception status flags will not be read and that floating point exceptions will be masked.

  • maytrap The compiler avoids transformations that may raise exceptions that would not have been raised by the original code. Constant folding performed by the compiler is exempt from this option.

  • strict The compiler ensures that all transformations strictly preserve the floating point exception semantics of the original code.

-ffp-eval-method=<value>

Specify the floating-point evaluation method for intermediate results within a single expression of the code.

Valid values are: source, double, and extended. For 64-bit targets, the default value is source. For 32-bit x86 targets however, in the case of NETBSD 6.99.26 and under, the default value is double; in the case of NETBSD greater than 6.99.26, with NoSSE, the default value is extended, with SSE the default value is source. Details:

  • source The compiler uses the floating-point type declared in the source program as the evaluation method.

  • double The compiler uses double as the floating-point evaluation method for all float expressions of type that is narrower than double.

  • extended The compiler uses long double as the floating-point evaluation method for all float expressions of type that is narrower than long double.

-f[no-]protect-parens

This option pertains to floating-point types, complex types with floating-point components, and vectors of these types. Some arithmetic expression transformations that are mathematically correct and permissible according to the C and C++ language standards may be incorrect when dealing with floating-point types, such as reassociation and distribution. Further, the optimizer may ignore parentheses when computing arithmetic expressions in circumstances where the parenthesized and unparenthesized expression express the same mathematical value. For example (a+b)+c is the same mathematical value as a+(b+c), but the optimizer is free to evaluate the additions in any order regardless of the parentheses. When enabled, this option forces the optimizer to honor the order of operations with respect to parentheses in all circumstances. Defaults to -fno-protect-parens.

Note that floating-point contraction (option -ffp-contract=) is disabled when -fprotect-parens is enabled. Also note that in safe floating-point modes, such as -ffp-model=precise or -ffp-model=strict, this option has no effect because the optimizer is prohibited from making unsafe transformations.

-fexcess-precision:

The C and C++ standards allow floating-point expressions to be computed as if intermediate results had more precision (and/or a wider range) than the type of the expression strictly allows. This is called excess precision arithmetic. Excess precision arithmetic can improve the accuracy of results (although not always), and it can make computation significantly faster if the target lacks direct hardware support for arithmetic in a particular type. However, it can also undermine strict floating-point reproducibility.

Under the standards, assignments and explicit casts force the operand to be converted to its formal type, discarding any excess precision. Because data can only flow between statements via an assignment, this means that the use of excess precision arithmetic is a reliable local property of a single statement, and results do not change based on optimization. However, when excess precision arithmetic is in use, Clang does not guarantee strict reproducibility, and future compiler releases may recognize more opportunities to use excess precision arithmetic, e.g. with floating-point builtins.

Clang does not use excess precision arithmetic for most types or on most targets. For example, even on pre-SSE X86 targets where float and double computations must be performed in the 80-bit X87 format, Clang rounds all intermediate results correctly for their type. Clang currently uses excess precision arithmetic by default only for the following types and targets:

  • _Float16 on X86 targets without AVX512-FP16.

The -fexcess-precision=<value> option can be used to control the use of excess precision arithmetic. Valid values are:

  • standard - The default. Allow the use of excess precision arithmetic under the constraints of the C and C++ standards. Has no effect except on the types and targets listed above.

  • fast - Accepted for GCC compatibility, but currently treated as an alias for standard.

  • 16 - Forces _Float16 operations to be emitted without using excess precision arithmetic.

-fcomplex-arithmetic=<value>:

This option specifies the implementation for complex multiplication and division.

Valid values are: basic, improved, full and promoted.

  • basic Implementation of complex division and multiplication using algebraic formulas at source precision. No special handling to avoid overflow. NaN and infinite values are not handled.

  • improved Implementation of complex division using the Smith algorithm at source precision. Smith’s algorithm for complex division. See SMITH, R. L. Algorithm 116: Complex division. Commun. ACM 5, 8 (1962). This value offers improved handling for overflow in intermediate calculations, but overflow may occur. NaN and infinite values are not handled in some cases.

  • full Implementation of complex division and multiplication using a call to runtime library functions (generally the case, but the BE might sometimes replace the library call if it knows enough about the potential range of the inputs). Overflow and non-finite values are handled by the library implementation. For the case of multiplication overflow will occur in accordance with normal floating-point rules. This is the default value.

  • promoted Implementation of complex division using algebraic formulas at higher precision. Overflow is handled. Non-finite values are handled in some cases. If the target does not have native support for a higher precision data type, the implementation for the complex operation using the Smith algorithm will be used. Overflow may still occur in some cases. NaN and infinite values are not handled.

-fcx-limited-range:

This option is aliased to -fcomplex-arithmetic=basic. It enables the naive mathematical formulas for complex division and multiplication with no NaN checking of results. The default is -fno-cx-limited-range aliased to -fcomplex-arithmetic=full. This option is enabled by the -ffast-math option.

-fcx-fortran-rules:

This option is aliased to -fcomplex-arithmetic=improved. It enables the naive mathematical formulas for complex multiplication and enables application of Smith’s algorithm for complex division. See SMITH, R. L. Algorithm 116: Complex division. Commun. ACM 5, 8 (1962). The default is -fno-cx-fortran-rules aliased to -fcomplex-arithmetic=full.

Accessing the floating point environment

Many targets allow floating point operations to be configured to control things such as how inexact results should be rounded and how exceptional conditions should be handled. This configuration is called the floating point environment. C and C++ restrict access to the floating point environment by default, and the compiler is allowed to assume that all operations are performed in the default environment. When code is compiled in this default mode, operations that depend on the environment (such as floating-point arithmetic and FLT_ROUNDS) may have undefined behavior if the dynamic environment is not the default environment; for example, FLT_ROUNDS may or may not simply return its default value for the target instead of reading the dynamic environment, and floating-point operations may be optimized as if the dynamic environment were the default. Similarly, it is undefined behavior to change the floating point environment in this default mode, for example by calling the fesetround function. C provides two pragmas to allow code to dynamically modify the floating point environment:

  • #pragma STDC FENV_ACCESS ON allows dynamic changes to the entire floating point environment.

  • #pragma STDC FENV_ROUND FE_DYNAMIC allows dynamic changes to just the floating point rounding mode. This may be more optimizable than FENV_ACCESS ON because the compiler can still ignore the possibility of floating-point exceptions by default.

Both of these can be used either at the start of a block scope, in which case they cover all code in that scope (unless they’re turned off in a child scope), or at the top level in a file, in which case they cover all subsequent function bodies until they’re turned off. Note that it is undefined behavior to enter code that is not covered by one of these pragmas from code that is covered by one of these pragmas unless the floating point environment has been restored to its default state. See the C standard for more information about these pragmas.

The command line option -frounding-math behaves as if the translation unit began with #pragma STDC FENV_ROUND FE_DYNAMIC. The command line option -ffp-model=strict behaves as if the translation unit began with #pragma STDC FENV_ACCESS ON.

Code that just wants to use a specific rounding mode for specific floating point operations can avoid most of the hazards of the dynamic floating point environment by using #pragma STDC FENV_ROUND with a value other than FE_DYNAMIC.

A note about crtfastmath.o

-ffast-math and -funsafe-math-optimizations without the -shared option cause crtfastmath.o to be automatically linked, which adds a static constructor that sets the FTZ/DAZ bits in MXCSR, affecting not only the current compilation unit but all static and shared libraries included in the program. This decision can be overridden by using either the flag -mdaz-ftz or -mno-daz-ftz to respectively link or not link crtfastmath.o.

A note about __FLT_EVAL_METHOD__

The __FLT_EVAL_METHOD__ is not defined as a traditional macro, and so it will not appear when dumping preprocessor macros. Instead, the value __FLT_EVAL_METHOD__ expands to is determined at the point of expansion either from the value set by the -ffp-eval-method command line option or from the target. This is because the __FLT_EVAL_METHOD__ macro cannot expand to the correct evaluation method in the presence of a #pragma which alters the evaluation method. An error is issued if __FLT_EVAL_METHOD__ is expanded inside a scope modified by #pragma clang fp eval_method.

A note about Floating Point Constant Evaluation

In C, the only place floating point operations are guaranteed to be evaluated during translation is in the initializers of variables of static storage duration, which are all notionally initialized before the program begins executing (and thus before a non-default floating point environment can be entered). But C++ has many more contexts where floating point constant evaluation occurs. Specifically: for static/thread-local variables, first try evaluating the initializer in a constant context, including in the constant floating point environment (just like in C), and then, if that fails, fall back to emitting runtime code to perform the initialization (which might in general be in a different floating point environment).

Consider this example when compiled with -frounding-math

constexpr float func_01(float x, float y) {
  return x + y;
}
float V1 = func_01(1.0F, 0x0.000001p0F);

The C++ rule is that initializers for static storage duration variables are first evaluated during translation (therefore, in the default rounding mode), and only evaluated at runtime (and therefore in the runtime rounding mode) if the compile-time evaluation fails. This is in line with the C rules; C11 F.8.5 says: All computation for automatic initialization is done (as if) at execution time; thus, it is affected by any operative modes and raises floating-point exceptions as required by IEC 60559 (provided the state for the FENV_ACCESS pragma is ‘‘on’’). All computation for initialization of objects that have static or thread storage duration is done (as if) at translation time. C++ generalizes this by adding another phase of initialization (at runtime) if the translation-time initialization fails, but the translation-time evaluation of the initializer of succeeds, it will be treated as a constant initializer.

Controlling Code Generation

Clang provides a number of ways to control code generation. The options are listed below.

-f[no-]sanitize=check1,check2,...

Turn on runtime checks for various forms of undefined or suspicious behavior.

This option controls whether Clang adds runtime checks for various forms of undefined or suspicious behavior, and is disabled by default. If a check fails, a diagnostic message is produced at runtime explaining the problem. The main checks are:

  • -fsanitize=address: AddressSanitizer, a memory error detector.

  • -fsanitize=thread: ThreadSanitizer, a data race detector.

  • -fsanitize=memory: MemorySanitizer, a detector of uninitialized reads. Requires instrumentation of all program code.

  • -fsanitize=undefined: UndefinedBehaviorSanitizer, a fast and compatible undefined behavior checker.

  • -fsanitize=dataflow: DataFlowSanitizer, a general data flow analysis.

  • -fsanitize=cfi: control flow integrity checks. Requires -flto.

  • -fsanitize=kcfi: kernel indirect call forward-edge control flow integrity.

  • -fsanitize=safe-stack: safe stack protection against stack-based memory corruption errors.

  • -fsanitize=realtime: RealtimeSanitizer, a real-time safety checker.

There are more fine-grained checks available: see the list of specific kinds of undefined behavior that can be detected and the list of control flow integrity schemes.

The -fsanitize= argument must also be provided when linking, in order to link to the appropriate runtime library.

It is not possible to combine more than one of the -fsanitize=address, -fsanitize=thread, and -fsanitize=memory checkers in the same program.

-f[no-]sanitize-recover=check1,check2,...
-f[no-]sanitize-recover[=all]

Controls which checks enabled by -fsanitize= flag are non-fatal. If the check is fatal, program will halt after the first error of this kind is detected and error report is printed.

By default, non-fatal checks are those enabled by UndefinedBehaviorSanitizer, except for -fsanitize=return and -fsanitize=unreachable. Some sanitizers may not support recovery (or not support it by default e.g. AddressSanitizer), and always crash the program after the issue is detected.

Note that the -fsanitize-trap flag has precedence over this flag. This means that if a check has been configured to trap elsewhere on the command line, or if the check traps by default, this flag will not have any effect unless that sanitizer’s trapping behavior is disabled with -fno-sanitize-trap.

For example, if a command line contains the flags -fsanitize=undefined -fsanitize-trap=undefined, the flag -fsanitize-recover=alignment will have no effect on its own; it will need to be accompanied by -fno-sanitize-trap=alignment.

-f[no-]sanitize-trap=check1,check2,...
-f[no-]sanitize-trap[=all]

Controls which checks enabled by the -fsanitize= flag trap. This option is intended for use in cases where the sanitizer runtime cannot be used (for instance, when building libc or a kernel module), or where the binary size increase caused by the sanitizer runtime is a concern.

This flag is only compatible with control flow integrity schemes and UndefinedBehaviorSanitizer checks other than vptr.

This flag is enabled by default for sanitizers in the cfi group.

-fsanitize-ignorelist=/path/to/ignorelist/file

Disable or modify sanitizer checks for objects (source files, functions, variables, types) listed in the file. See Sanitizer special case list for file format description.

-fno-sanitize-ignorelist

Don’t use ignorelist file, if it was specified earlier in the command line.

-f[no-]sanitize-coverage=[type,features,...]

Enable simple code coverage in addition to certain sanitizers. See SanitizerCoverage for more details.

-f[no-]sanitize-address-outline-instrumentation

Controls how address sanitizer code is generated. If enabled will always use a function call instead of inlining the code. Turning this option on could reduce the binary size, but might result in a worse run-time performance.

See :doc: AddressSanitizer for more details.

-f[no-]sanitize-stats

Enable simple statistics gathering for the enabled sanitizers. See SanitizerStats for more details.

-fsanitize-undefined-trap-on-error

Deprecated alias for -fsanitize-trap=undefined.

-fsanitize-cfi-cross-dso

Enable cross-DSO control flow integrity checks. This flag modifies the behavior of sanitizers in the cfi group to allow checking of cross-DSO virtual and indirect calls.

-fsanitize-cfi-icall-generalize-pointers

Generalize pointers in return and argument types in function type signatures checked by Control Flow Integrity indirect call checking. See Control Flow Integrity for more details.

-fsanitize-cfi-icall-experimental-normalize-integers

Normalize integers in return and argument types in function type signatures checked by Control Flow Integrity indirect call checking. See Control Flow Integrity for more details.

This option is currently experimental.

-fstrict-vtable-pointers

Enable optimizations based on the strict rules for overwriting polymorphic C++ objects, i.e. the vptr is invariant during an object’s lifetime. This enables better devirtualization. Turned off by default, because it is still experimental.

-fwhole-program-vtables

Enable whole-program vtable optimizations, such as single-implementation devirtualization and virtual constant propagation, for classes with hidden LTO visibility. Requires -flto.

-f[no]split-lto-unit

Controls splitting the LTO unit into regular LTO and ThinLTO portions, when compiling with -flto=thin. Defaults to false unless -fsanitize=cfi or -fwhole-program-vtables are specified, in which case it defaults to true. Splitting is required with fsanitize=cfi, and it is an error to disable via -fno-split-lto-unit. Splitting is optional with -fwhole-program-vtables, however, it enables more aggressive whole program vtable optimizations (specifically virtual constant propagation).

When enabled, vtable definitions and select virtual functions are placed in the split regular LTO module, enabling more aggressive whole program vtable optimizations required for CFI and virtual constant propagation. However, this can increase the LTO link time and memory requirements over pure ThinLTO, as all split regular LTO modules are merged and LTO linked with regular LTO.

-fforce-emit-vtables

In order to improve devirtualization, forces emitting of vtables even in modules where it isn’t necessary. It causes more inline virtual functions to be emitted.

-fno-assume-sane-operator-new

Don’t assume that the C++’s new operator is sane.

This option tells the compiler to do not assume that C++’s global new operator will always return a pointer that does not alias any other pointer when the function returns.

-fassume-nothrow-exception-dtor

Assume that an exception object’ destructor will not throw, and generate less code for catch handlers. A throw expression of a type with a potentially-throwing destructor will lead to an error.

By default, Clang assumes that the exception object may have a throwing destructor. For the Itanium C++ ABI, Clang generates a landing pad to destroy local variables and call _Unwind_Resume for the code catch (...) { ... }. This option tells Clang that an exception object’s destructor will not throw and code simplification is possible.

-ftrap-function=[name]

Instruct code generator to emit a function call to the specified function name for __builtin_trap().

LLVM code generator translates __builtin_trap() to a trap instruction if it is supported by the target ISA. Otherwise, the builtin is translated into a call to abort. If this option is set, then the code generator will always lower the builtin to a call to the specified function regardless of whether the target ISA has a trap instruction. This option is useful for environments (e.g. deeply embedded) where a trap cannot be properly handled, or when some custom behavior is desired.

-ftls-model=[model]

Select which TLS model to use.

Valid values are: global-dynamic, local-dynamic, initial-exec and local-exec. The default value is global-dynamic. The compiler may use a different model if the selected model is not supported by the target, or if a more efficient model can be used. The TLS model can be overridden per variable using the tls_model attribute.

-femulated-tls

Select emulated TLS model, which overrides all -ftls-model choices.

In emulated TLS mode, all access to TLS variables are converted to calls to __emutls_get_address in the runtime library.

-mhwdiv=[values]

Select the ARM modes (arm or thumb) that support hardware division instructions.

Valid values are: arm, thumb and arm,thumb. This option is used to indicate which mode (arm or thumb) supports hardware division instructions. This only applies to the ARM architecture.

-m[no-]crc

Enable or disable CRC instructions.

This option is used to indicate whether CRC instructions are to be generated. This only applies to the ARM architecture.

CRC instructions are enabled by default on ARMv8.

-mgeneral-regs-only

Generate code which only uses the general purpose registers.

This option restricts the generated code to use general registers only. This only applies to the AArch64 architecture.

-mcompact-branches=[values]

Control the usage of compact branches for MIPSR6.

Valid values are: never, optimal and always. The default value is optimal which generates compact branches when a delay slot cannot be filled. never disables the usage of compact branches and always generates compact branches whenever possible.

-f[no-]max-type-align=[number]

Instruct the code generator to not enforce a higher alignment than the given number (of bytes) when accessing memory via an opaque pointer or reference. This cap is ignored when directly accessing a variable or when the pointee type has an explicit “aligned” attribute.

The value should usually be determined by the properties of the system allocator. Some builtin types, especially vector types, have very high natural alignments; when working with values of those types, Clang usually wants to use instructions that take advantage of that alignment. However, many system allocators do not promise to return memory that is more than 8-byte or 16-byte-aligned. Use this option to limit the alignment that the compiler can assume for an arbitrary pointer, which may point onto the heap.

This option does not affect the ABI alignment of types; the layout of structs and unions and the value returned by the alignof operator remain the same.

This option can be overridden on a case-by-case basis by putting an explicit “aligned” alignment on a struct, union, or typedef. For example:

#include <immintrin.h>
// Make an aligned typedef of the AVX-512 16-int vector type.
typedef __v16si __aligned_v16si __attribute__((aligned(64)));

void initialize_vector(__aligned_v16si *v) {
  // The compiler may assume that ‘v’ is 64-byte aligned, regardless of the
  // value of -fmax-type-align.
}
-faddrsig, -fno-addrsig

Controls whether Clang emits an address-significance table into the object file. Address-significance tables allow linkers to implement safe ICF without the false positives that can result from other implementation techniques such as relocation scanning. Address-significance tables are enabled by default on ELF targets when using the integrated assembler. This flag currently only has an effect on ELF targets.

-f[no]-unique-internal-linkage-names

Controls whether Clang emits a unique (best-effort) symbol name for internal linkage symbols. When this option is set, compiler hashes the main source file path from the command line and appends it to all internal symbols. If a program contains multiple objects compiled with the same command-line source file path, the symbols are not guaranteed to be unique. This option is particularly useful in attributing profile information to the correct function when multiple functions with the same private linkage name exist in the binary.

It should be noted that this option cannot guarantee uniqueness and the following is an example where it is not unique when two modules contain symbols with the same private linkage name:

$ cd $P/foo && clang -c -funique-internal-linkage-names name_conflict.c
$ cd $P/bar && clang -c -funique-internal-linkage-names name_conflict.c
$ cd $P && clang foo/name_conflict.o && bar/name_conflict.o
-f[no]-basic-block-address-map:
Emits a ``SHT_LLVM_BB_ADDR_MAP`` section which includes address offsets for each
basic block in the program, relative to the parent function address.
-fbasic-block-sections=[all, list=<arg>, none]

Controls how Clang emits text sections for basic blocks. With values all and list=<arg>, each basic block or a subset of basic blocks can be placed in its own unique section.

With the list=<arg> option, a file containing the subset of basic blocks that need to placed in unique sections can be specified. The format of the file is as follows. For example, list=spec.txt where spec.txt is the following:

!foo
!!2
!_Z3barv

will place the machine basic block with id 2 in function foo in a unique section. It will also place all basic blocks of functions bar in unique sections.

Further, section clusters can also be specified using the list=<arg> option. For example, list=spec.txt where spec.txt contains:

!foo
!!1 !!3 !!5
!!2 !!4 !!6

will create two unique sections for function foo with the first containing the odd numbered basic blocks and the second containing the even numbered basic blocks.

Basic block sections allow the linker to reorder basic blocks and enables link-time optimizations like whole program inter-procedural basic block reordering.

-fcodegen-data-generate[=<path>]

Emit the raw codegen (CG) data into custom sections in the object file. Currently, this option also combines the raw CG data from the object files into an indexed CG data file specified by the <path>, for LLD MachO only. When the <path> is not specified, default.cgdata is created. The CG data file combines all the outlining instances that occurred locally in each object file.

$ clang -fuse-ld=lld -Oz -fcodegen-data-generate code.cc

For linkers that do not yet support this feature, llvm-cgdata can be used manually to merge this CG data in object files.

$ clang -c -fuse-ld=lld -Oz -fcodegen-data-generate code.cc
$ llvm-cgdata --merge -o default.cgdata code.o
-fcodegen-data-use[=<path>]

Read the codegen data from the specified path to more effectively outline functions across compilation units. When the <path> is not specified, default.cgdata is used. This option can create many identically outlined functions that can be optimized by the conventional linker’s identical code folding (ICF).

$ clang -fuse-ld=lld -Oz -Wl,--icf=safe -fcodegen-data-use code.cc

Profile Guided Optimization

Profile information enables better optimization. For example, knowing that a branch is taken very frequently helps the compiler make better decisions when ordering basic blocks. Knowing that a function foo is called more frequently than another function bar helps the inliner. Optimization levels -O2 and above are recommended for use of profile guided optimization.

Clang supports profile guided optimization with two different kinds of profiling. A sampling profiler can generate a profile with very low runtime overhead, or you can build an instrumented version of the code that collects more detailed profile information. Both kinds of profiles can provide execution counts for instructions in the code and information on branches taken and function invocation.

Regardless of which kind of profiling you use, be careful to collect profiles by running your code with inputs that are representative of the typical behavior. Code that is not exercised in the profile will be optimized as if it is unimportant, and the compiler may make poor optimization choices for code that is disproportionately used while profiling.

Differences Between Sampling and Instrumentation

Although both techniques are used for similar purposes, there are important differences between the two:

  1. Profile data generated with one cannot be used by the other, and there is no conversion tool that can convert one to the other. So, a profile generated via -fprofile-generate or -fprofile-instr-generate must be used with -fprofile-use or -fprofile-instr-use. Similarly, sampling profiles generated by external profilers must be converted and used with -fprofile-sample-use or -fauto-profile.

  2. Instrumentation profile data can be used for code coverage analysis and optimization.

  3. Sampling profiles can only be used for optimization. They cannot be used for code coverage analysis. Although it would be technically possible to use sampling profiles for code coverage, sample-based profiles are too coarse-grained for code coverage purposes; it would yield poor results.

  4. Sampling profiles must be generated by an external tool. The profile generated by that tool must then be converted into a format that can be read by LLVM. The section on sampling profilers describes one of the supported sampling profile formats.

Using Sampling Profilers

Sampling profilers are used to collect runtime information, such as hardware counters, while your application executes. They are typically very efficient and do not incur a large runtime overhead. The sample data collected by the profiler can be used during compilation to determine what the most executed areas of the code are.

Using the data from a sample profiler requires some changes in the way a program is built. Before the compiler can use profiling information, the code needs to execute under the profiler. The following is the usual build cycle when using sample profilers for optimization:

  1. Build the code with source line table information. You can use all the usual build flags that you always build your application with. The only requirement is that DWARF debug info including source line information is generated. This DWARF information is important for the profiler to be able to map instructions back to source line locations. The usefulness of this DWARF information can be improved with the -fdebug-info-for-profiling and -funique-internal-linkage-names options.

    On Linux:

    $ clang++ -O2 -gline-tables-only \
  -fdebug-info-for-profiling -funique-internal-linkage-names \
  code.cc -o code

    While MSVC-style targets default to CodeView debug information, DWARF debug information is required to generate source-level LLVM profiles. Use -gdwarf to include DWARF debug information:

    > clang-cl /O2 -gdwarf -gline-tables-only ^
  /clang:-fdebug-info-for-profiling /clang:-funique-internal-linkage-names ^
  code.cc /Fe:code /fuse-ld=lld /link /debug:dwarf

Note

-funique-internal-linkage-names generates unique names based on given command-line source file paths. If your build system uses absolute source paths and these paths may change between steps 1 and 4, then the uniqued function names may change and result in unused profile data. Consider omitting this option in such cases.

  1. Run the executable under a sampling profiler. The specific profiler you use does not really matter, as long as its output can be converted into the format that the LLVM optimizer understands.

    Two such profilers are the Linux Perf profiler (https://perf.wiki.kernel.org/) and Intel’s Sampling Enabling Product (SEP), available as part of Intel VTune. While Perf is Linux-specific, SEP can be used on Linux, Windows, and FreeBSD.

    The LLVM tool llvm-profgen can convert output of either Perf or SEP. An external project, AutoFDO, also provides a create_llvm_prof tool which supports Linux Perf output.

    When using Perf:

    $ perf record -b -e BR_INST_RETIRED.NEAR_TAKEN:uppp ./code

    If the event above is unavailable, branches:u is probably next-best.

    Note the use of the -b flag. This tells Perf to use the Last Branch Record (LBR) to record call chains. While this is not strictly required, it provides better call information, which improves the accuracy of the profile data.

    When using SEP:

    $ sep -start -out code.tb7 -ec BR_INST_RETIRED.NEAR_TAKEN:precise=yes:pdir -lbr no_filter:usr -perf-script brstack -app ./code

    This produces a code.perf.data.script output which can be used with llvm-profgen’s --perfscript input option.

  2. Convert the collected profile data to LLVM’s sample profile format. This is currently supported via the AutoFDO converter create_llvm_prof. Once built and installed, you can convert the perf.data file to LLVM using the command:

    $ create_llvm_prof --binary=./code --out=code.prof

    This will read perf.data and the binary file ./code and emit the profile data in code.prof. Note that if you ran perf without the -b flag, you need to use --use_lbr=false when calling create_llvm_prof.

    Alternatively, the LLVM tool llvm-profgen can also be used to generate the LLVM sample profile:

    $ llvm-profgen --binary=./code --output=code.prof --perfdata=perf.data

    When using SEP the output is in the textual format corresponding to llvm-profgen --perfscript. For example:

    $ llvm-profgen --binary=./code --output=code.prof --perfscript=code.perf.data.script

  3. Build the code again using the collected profile. This step feeds the profile back to the optimizers. This should result in a binary that executes faster than the original one. Note that you are not required to build the code with the exact same arguments that you used in the first step. The only requirement is that you build the code with the same debug info options and -fprofile-sample-use.

    On Linux:

    $ clang++ -O2 -gline-tables-only \
  -fdebug-info-for-profiling -funique-internal-linkage-names \
  -fprofile-sample-use=code.prof code.cc -o code

    On Windows:

    > clang-cl /O2 -gdwarf -gline-tables-only ^
  /clang:-fdebug-info-for-profiling /clang:-funique-internal-linkage-names ^
  /fprofile-sample-use=code.prof code.cc /Fe:code /fuse-ld=lld /link /debug:dwarf

    [OPTIONAL] Sampling-based profiles can have inaccuracies or missing block/ edge counters. The profile inference algorithm (profi) can be used to infer missing blocks and edge counts, and improve the quality of profile data. Enable it with -fsample-profile-use-profi. For example, on Linux:

    $ clang++ -fsample-profile-use-profi -O2 -gline-tables-only \
  -fdebug-info-for-profiling -funique-internal-linkage-names \
  -fprofile-sample-use=code.prof code.cc -o code

    On Windows:

    > clang-cl /clang:-fsample-profile-use-profi /O2 -gdwarf -gline-tables-only ^
  /clang:-fdebug-info-for-profiling /clang:-funique-internal-linkage-names ^
  /fprofile-sample-use=code.prof code.cc /Fe:code /fuse-ld=lld /link /debug:dwarf
Sample Profile Formats

Since external profilers generate profile data in a variety of custom formats, the data generated by the profiler must be converted into a format that can be read by the backend. LLVM supports three different sample profile formats:

  1. ASCII text. This is the easiest one to generate. The file is divided into sections, which correspond to each of the functions with profile information. The format is described below. It can also be generated from the binary or gcov formats using the llvm-profdata tool.

  2. Binary encoding. This uses a more efficient encoding that yields smaller profile files. This is the format generated by the create_llvm_prof tool in https://github.com/google/autofdo.

  3. GCC encoding. This is based on the gcov format, which is accepted by GCC. It is only interesting in environments where GCC and Clang co-exist. This encoding is only generated by the create_gcov tool in https://github.com/google/autofdo. It can be read by LLVM and llvm-profdata, but it cannot be generated by either.

If you are using Linux Perf to generate sampling profiles, you can use the conversion tool create_llvm_prof described in the previous section. Otherwise, you will need to write a conversion tool that converts your profiler’s native format into one of these three.

Sample Profile Text Format

This section describes the ASCII text format for sampling profiles. It is, arguably, the easiest one to generate. If you are interested in generating any of the other two, consult the ProfileData library in LLVM’s source tree (specifically, include/llvm/ProfileData/SampleProfReader.h).

function1:total_samples:total_head_samples
 offset1[.discriminator]: number_of_samples [fn1:num fn2:num ... ]
 offset2[.discriminator]: number_of_samples [fn3:num fn4:num ... ]
 ...
 offsetN[.discriminator]: number_of_samples [fn5:num fn6:num ... ]
 offsetA[.discriminator]: fnA:num_of_total_samples
  offsetA1[.discriminator]: number_of_samples [fn7:num fn8:num ... ]
  offsetA1[.discriminator]: number_of_samples [fn9:num fn10:num ... ]
  offsetB[.discriminator]: fnB:num_of_total_samples
   offsetB1[.discriminator]: number_of_samples [fn11:num fn12:num ... ]

This is a nested tree in which the indentation represents the nesting level of the inline stack. There are no blank lines in the file. And the spacing within a single line is fixed. Additional spaces will result in an error while reading the file.

Any line starting with the ‘#’ character is completely ignored.

Inlined calls are represented with indentation. The Inline stack is a stack of source locations in which the top of the stack represents the leaf function, and the bottom of the stack represents the actual symbol to which the instruction belongs.

Function names must be mangled in order for the profile loader to match them in the current translation unit. The two numbers in the function header specify how many total samples were accumulated in the function (first number), and the total number of samples accumulated in the prologue of the function (second number). This head sample count provides an indicator of how frequently the function is invoked.

There are two types of lines in the function body.

  • Sampled line represents the profile information of a source location. offsetN[.discriminator]: number_of_samples [fn5:num fn6:num ... ]

  • Callsite line represents the profile information of an inlined callsite. offsetA[.discriminator]: fnA:num_of_total_samples

Each sampled line may contain several items. Some are optional (marked below):

  1. Source line offset. This number represents the line number in the function where the sample was collected. The line number is always relative to the line where symbol of the function is defined. So, if the function has its header at line 280, the offset 13 is at line 293 in the file.

    Note that this offset should never be a negative number. This could happen in cases like macros. The debug machinery will register the line number at the point of macro expansion. So, if the macro was expanded in a line before the start of the function, the profile converter should emit a 0 as the offset (this means that the optimizers will not be able to associate a meaningful weight to the instructions in the macro).

  2. [OPTIONAL] Discriminator. This is used if the sampled program was compiled with DWARF discriminator support (http://wiki.dwarfstd.org/index.php?title=Path_Discriminators). DWARF discriminators are unsigned integer values that allow the compiler to distinguish between multiple execution paths on the same source line location.

    For example, consider the line of code if (cond) foo(); else bar();. If the predicate cond is true 80% of the time, then the edge into function foo should be considered to be taken most of the time. But both calls to foo and bar are at the same source line, so a sample count at that line is not sufficient. The compiler needs to know which part of that line is taken more frequently.

    This is what discriminators provide. In this case, the calls to foo and bar will be at the same line, but will have different discriminator values. This allows the compiler to correctly set edge weights into foo and bar.

  3. Number of samples. This is an integer quantity representing the number of samples collected by the profiler at this source location.

  4. [OPTIONAL] Potential call targets and samples. If present, this line contains a call instruction. This models both direct and number of samples. For example,

    130: 7  foo:3  bar:2  baz:7

    The above means that at relative line offset 130 there is a call instruction that calls one of foo(), bar() and baz(), with baz() being the relatively more frequently called target.

As an example, consider a program with the call chain main -> foo -> bar. When built with optimizations enabled, the compiler may inline the calls to bar and foo inside main. The generated profile could then be something like this:

main:35504:0
1: _Z3foov:35504
  2: _Z32bari:31977
  1.1: 31977
2: 0

This profile indicates that there were a total of 35,504 samples collected in main. All of those were at line 1 (the call to foo). Of those, 31,977 were spent inside the body of bar. The last line of the profile (2: 0) corresponds to line 2 inside main. No samples were collected there.

Profiling with Instrumentation

Clang also supports profiling via instrumentation. This requires building a special instrumented version of the code and has some runtime overhead during the profiling, but it provides more detailed results than a sampling profiler. It also provides reproducible results, at least to the extent that the code behaves consistently across runs.

Clang supports two types of instrumentation: frontend-based and IR-based. Frontend-based instrumentation can be enabled with the option -fprofile-instr-generate, and IR-based instrumentation can be enabled with the option -fprofile-generate. For best performance with PGO, IR-based instrumentation should be used. It has the benefits of lower instrumentation overhead, smaller raw profile size, and better runtime performance. Frontend-based instrumentation, on the other hand, has better source correlation, so it should be used with source line-based coverage testing.

The flag -fcs-profile-generate also instruments programs using the same instrumentation method as -fprofile-generate. However, it performs a post-inline late instrumentation and can produce context-sensitive profiles.

Here are the steps for using profile guided optimization with instrumentation:

  1. Build an instrumented version of the code by compiling and linking with the -fprofile-generate or -fprofile-instr-generate option.

    $ clang++ -O2 -fprofile-instr-generate code.cc -o code

  2. Run the instrumented executable with inputs that reflect the typical usage. By default, the profile data will be written to a default.profraw file in the current directory. You can override that default by using option -fprofile-instr-generate= or by setting the LLVM_PROFILE_FILE environment variable to specify an alternate file. If non-default file name is specified by both the environment variable and the command line option, the environment variable takes precedence. The file name pattern specified can include different modifiers: %p, %h, %m, %t, and %c.

    Any instance of %p in that file name will be replaced by the process ID, so that you can easily distinguish the profile output from multiple runs.

    $ LLVM_PROFILE_FILE="code-%p.profraw" ./code

    The modifier %h can be used in scenarios where the same instrumented binary is run in multiple different host machines dumping profile data to a shared network based storage. The %h specifier will be substituted with the hostname so that profiles collected from different hosts do not clobber each other.

    While the use of %p specifier can reduce the likelihood for the profiles dumped from different processes to clobber each other, such clobbering can still happen because of the pid re-use by the OS. Another side-effect of using %p is that the storage requirement for raw profile data files is greatly increased. To avoid issues like this, the %m specifier can used in the profile name. When this specifier is used, the profiler runtime will substitute %m with a unique integer identifier associated with the instrumented binary. Additionally, multiple raw profiles dumped from different processes that share a file system (can be on different hosts) will be automatically merged by the profiler runtime during the dumping. If the program links in multiple instrumented shared libraries, each library will dump the profile data into its own profile data file (with its unique integer id embedded in the profile name). Note that the merging enabled by %m is for raw profile data generated by profiler runtime. The resulting merged “raw” profile data file still needs to be converted to a different format expected by the compiler ( see step 3 below).

    $ LLVM_PROFILE_FILE="code-%m.profraw" ./code

    See this section about the %t, and %c modifiers.

  3. Combine profiles from multiple runs and convert the “raw” profile format to the input expected by clang. Use the merge command of the llvm-profdata tool to do this.

    $ llvm-profdata merge -output=code.profdata code-*.profraw

    Note that this step is necessary even when there is only one “raw” profile, since the merge operation also changes the file format.

  4. Build the code again using the -fprofile-use or -fprofile-instr-use option to specify the collected profile data.

    $ clang++ -O2 -fprofile-instr-use=code.profdata code.cc -o code

    You can repeat step 4 as often as you like without regenerating the profile. As you make changes to your code, clang may no longer be able to use the profile data. It will warn you when this happens.

Note that -fprofile-use option is semantically equivalent to its GCC counterpart, it does not handle profile formats produced by GCC. Both -fprofile-use and -fprofile-instr-use accept profiles in the indexed format, regardeless whether it is produced by frontend or the IR pass.

-fprofile-generate[=<dirname>]

The -fprofile-generate and -fprofile-generate= flags will use an alternative instrumentation method for profile generation. When given a directory name, it generates the profile file default_%m.profraw in the directory named dirname if specified. If dirname does not exist, it will be created at runtime. %m specifier will be substituted with a unique id documented in step 2 above. In other words, with -fprofile-generate[=<dirname>] option, the “raw” profile data automatic merging is turned on by default, so there will no longer any risk of profile clobbering from different running processes. For example,

$ clang++ -O2 -fprofile-generate=yyy/zzz code.cc -o code

When code is executed, the profile will be written to the file yyy/zzz/default_xxxx.profraw.

To generate the profile data file with the compiler readable format, the llvm-profdata tool can be used with the profile directory as the input:

$ llvm-profdata merge -output=code.profdata yyy/zzz/

If the user wants to turn off the auto-merging feature, or simply override the the profile dumping path specified at command line, the environment variable LLVM_PROFILE_FILE can still be used to override the directory and filename for the profile file at runtime. To override the path and filename at compile time, use -Xclang -fprofile-instrument-path=/path/to/file_pattern.profraw.

-fcs-profile-generate[=<dirname>]

The -fcs-profile-generate and -fcs-profile-generate= flags will use the same instrumentation method, and generate the same profile as in the -fprofile-generate and -fprofile-generate= flags. The difference is that the instrumentation is performed after inlining so that the resulted profile has a better context sensitive information. They cannot be used together with -fprofile-generate and -fprofile-generate= flags. They are typically used in conjunction with -fprofile-use flag. The profile generated by -fcs-profile-generate and -fprofile-generate can be merged by llvm-profdata. A use example:

$ clang++ -O2 -fprofile-generate=yyy/zzz code.cc -o code
$ ./code
$ llvm-profdata merge -output=code.profdata yyy/zzz/

The first few steps are the same as that in -fprofile-generate compilation. Then perform a second round of instrumentation.

$ clang++ -O2 -fprofile-use=code.profdata -fcs-profile-generate=sss/ttt \
  -o cs_code
$ ./cs_code
$ llvm-profdata merge -output=cs_code.profdata sss/ttt code.profdata

The resulted cs_code.prodata combines code.profdata and the profile generated from binary cs_code. Profile cs_code.profata can be used by -fprofile-use compilation.

$ clang++ -O2 -fprofile-use=cs_code.profdata

The above command will read both profiles to the compiler at the identical point of instrumentations.

-fprofile-use[=<pathname>]

Without any other arguments, -fprofile-use behaves identically to -fprofile-instr-use. Otherwise, if pathname is the full path to a profile file, it reads from that file. If pathname is a directory name, it reads from pathname/default.profdata.

-fprofile-update[=<method>]

Unless -fsanitize=thread is specified, the default is single, which uses non-atomic increments. The counters can be inaccurate under thread contention. atomic uses atomic increments which is accurate but has overhead. prefer-atomic will be transformed to atomic when supported by the target, or single otherwise.

Fine Tuning Profile Collection

The PGO infrastructure provides user program knobs to fine tune profile collection. Specifically, the PGO runtime provides the following functions that can be used to control the regions in the program where profiles should be collected.

  • void __llvm_profile_set_filename(const char *Name): changes the name of the profile file to Name.

  • void __llvm_profile_reset_counters(void): resets all counters to zero.

  • int __llvm_profile_dump(void): write the profile data to disk.

  • int __llvm_orderfile_dump(void): write the order file to disk.

For example, the following pattern can be used to skip profiling program initialization, profile two specific hot regions, and skip profiling program cleanup:

int main() {
  initialize();

  // Reset all profile counters to 0 to omit profile collected during
  // initialize()'s execution.
  __llvm_profile_reset_counters();
  ... hot region 1
  // Dump the profile for hot region 1.
  __llvm_profile_set_filename("region1.profraw");
  __llvm_profile_dump();

  // Reset counters before proceeding to hot region 2.
  __llvm_profile_reset_counters();
  ... hot region 2
  // Dump the profile for hot region 2.
  __llvm_profile_set_filename("region2.profraw");
  __llvm_profile_dump();

  // Since the profile has been dumped, no further profile data
  // will be collected beyond the above __llvm_profile_dump().
  cleanup();
  return 0;
}

These APIs’ names can be introduced to user programs in two ways. They can be declared as weak symbols on platforms which support treating weak symbols as null during linking. For example, the user can have

__attribute__((weak)) int __llvm_profile_dump(void);

// Then later in the same source file
if (__llvm_profile_dump)
  if (__llvm_profile_dump() != 0) { ... }
// The first if condition tests if the symbol is actually defined.
// Profile dumping only happens if the symbol is defined. Hence,
// the user program works correctly during normal (not profile-generate)
// executions.

Alternatively, the user program can include the header profile/instr_prof_interface.h, which contains the API names. For example,

#include "profile/instr_prof_interface.h"

// Then later in the same source file
if (__llvm_profile_dump() != 0) { ... }

The user code does not need to check if the API names are defined, because these names are automatically replaced by (0) or the equivalence of noop if the clang is not compiling for profile generation.

Such replacement can happen because clang adds one of two macros depending on the -fprofile-generate and the -fprofile-use flags.

  • __LLVM_INSTR_PROFILE_GENERATE: defined when one of -fprofile[-instr]-generate/-fcs-profile-generate is in effect.

  • __LLVM_INSTR_PROFILE_USE: defined when one of -fprofile-use/-fprofile-instr-use is in effect.

The two macros can be used to provide more flexibiilty so a user program can execute code specifically intended for profile generate or profile use. For example, a user program can have special logging during profile generate:

#if __LLVM_INSTR_PROFILE_GENERATE
expensive_logging_of_full_program_state();
#endif

The logging is automatically excluded during a normal build of the program, hence it does not impact performance during a normal execution.

It is advised to use such fine tuning only in a program’s cold regions. The weak symbols can introduce extra control flow (the if checks), while the macros (hence declarations they guard in profile/instr_prof_interface.h) can change the control flow of the functions that use them between profile generation and profile use (which can lead to discarded counters in such functions). Using these APIs in the program’s cold regions introduces less overhead and leads to more optimized code.

Disabling Instrumentation

In certain situations, it may be useful to disable profile generation or use for specific files in a build, without affecting the main compilation flags used for the other files in the project.

In these cases, you can use the flag -fno-profile-instr-generate (or -fno-profile-generate) to disable profile generation, and -fno-profile-instr-use (or -fno-profile-use) to disable profile use.

Note that these flags should appear after the corresponding profile flags to have an effect.

Note

When none of the translation units inside a binary is instrumented, in the case of Fuchsia the profile runtime will not be linked into the binary and no profile will be produced, while on other platforms the profile runtime will be linked and profile will be produced but there will not be any counters.

Instrumenting only selected files or functions

Sometimes it’s useful to only instrument certain files or functions. For example in automated testing infrastructure, it may be desirable to only instrument files or functions that were modified by a patch to reduce the overhead of instrumenting a full system.

This can be done using the -fprofile-list option.

-fprofile-list=<pathname>

This option can be used to apply profile instrumentation only to selected files or functions. pathname should point to a file in the Sanitizer special case list format which selects which files and functions to instrument.

$ clang++ -O2 -fprofile-instr-generate -fprofile-list=fun.list code.cc -o code

The option can be specified multiple times to pass multiple files.

$ clang++ -O2 -fprofile-instr-generate -fcoverage-mapping -fprofile-list=fun.list -fprofile-list=code.list code.cc -o code

Supported sections are [clang], [llvm], and [csllvm] representing clang PGO, IRPGO, and CSIRPGO, respectively. Supported prefixes are function and source. Supported categories are allow, skip, and forbid. skip adds the skipprofile attribute while forbid adds the noprofile attribute to the appropriate function. Use default:<allow|skip|forbid> to specify the default category.

$ cat fun.list
# The following cases are for clang instrumentation.
[clang]

# We might not want to profile functions that are inlined in many places.
function:inlinedLots=skip

# We want to forbid profiling where it might be dangerous.
source:lib/unsafe/*.cc=forbid

# Otherwise we allow profiling.
default:allow
Older Prefixes

An older format is also supported, but it is only able to add the noprofile attribute. To filter individual functions or entire source files use fun:<name> or src:<file> respectively. To exclude a function or a source file, use !fun:<name> or !src:<file> respectively. The format also supports wildcard expansion. The compiler generated functions are assumed to be located in the main source file. It is also possible to restrict the filter to a particular instrumentation type by using a named section.

# all functions whose name starts with foo will be instrumented.
fun:foo*

# except for foo1 which will be excluded from instrumentation.
!fun:foo1

# every function in path/to/foo.cc will be instrumented.
src:path/to/foo.cc

# bar will be instrumented only when using backend instrumentation.
# Recognized section names are clang, llvm and csllvm.
[llvm]
fun:bar

When the file contains only excludes, all files and functions except for the excluded ones will be instrumented. Otherwise, only the files and functions specified will be instrumented.

Instrument function groups

Sometimes it is desirable to minimize the size overhead of instrumented binaries. One way to do this is to partition functions into groups and only instrument functions in a specified group. This can be done using the -fprofile-function-groups and -fprofile-selected-function-group options.

-fprofile-function-groups=<N>, -fprofile-selected-function-group=<i>

The following uses 3 groups

$ clang++ -Oz -fprofile-generate=group_0/ -fprofile-function-groups=3 -fprofile-selected-function-group=0 code.cc -o code.0
$ clang++ -Oz -fprofile-generate=group_1/ -fprofile-function-groups=3 -fprofile-selected-function-group=1 code.cc -o code.1
$ clang++ -Oz -fprofile-generate=group_2/ -fprofile-function-groups=3 -fprofile-selected-function-group=2 code.cc -o code.2

After collecting raw profiles from the three binaries, they can be merged into a single profile like normal.

$ llvm-profdata merge -output=code.profdata group_*/*.profraw

Profile remapping

When the program is compiled after a change that affects many symbol names, pre-existing profile data may no longer match the program. For example:

  • switching from libstdc++ to libc++ will result in the mangled names of all functions taking standard library types to change

  • renaming a widely-used type in C++ will result in the mangled names of all functions that have parameters involving that type to change

  • moving from a 32-bit compilation to a 64-bit compilation may change the underlying type of size_t and similar types, resulting in changes to manglings

Clang allows use of a profile remapping file to specify that such differences in mangled names should be ignored when matching the profile data against the program.

-fprofile-remapping-file=<file>

Specifies a file containing profile remapping information, that will be used to match mangled names in the profile data to mangled names in the program.

The profile remapping file is a text file containing lines of the form

fragmentkind fragment1 fragment2

where fragmentkind is one of name, type, or encoding, indicating whether the following mangled name fragments are <name>s, <type>s, or <encoding>s, respectively. Blank lines and lines starting with # are ignored.

For convenience, built-in <substitution>s such as St and Ss are accepted as <name>s (even though they technically are not <name>s).

For example, to specify that absl::string_view and std::string_view should be treated as equivalent when matching profile data, the following remapping file could be used:

# absl::string_view is considered equivalent to std::string_view
type N4absl11string_viewE St17basic_string_viewIcSt11char_traitsIcEE

# std:: might be std::__1:: in libc++ or std::__cxx11:: in libstdc++
name 3std St3__1
name 3std St7__cxx11

Matching profile data using a profile remapping file is supported on a best-effort basis. For example, information regarding indirect call targets is currently not remapped. For best results, you are encouraged to generate new profile data matching the updated program, or to remap the profile data using the llvm-cxxmap and llvm-profdata merge tools.

Note

Profile data remapping is currently only supported for C++ mangled names following the Itanium C++ ABI mangling scheme. This covers all C++ targets supported by Clang other than Windows.

GCOV-based Profiling

GCOV is a test coverage program, it helps to know how often a line of code is executed. When instrumenting the code with --coverage option, some counters are added for each edge linking basic blocks.

At compile time, gcno files are generated containing information about blocks and edges between them. At runtime the counters are incremented and at exit the counters are dumped in gcda files.

The tool llvm-cov gcov will parse gcno, gcda and source files to generate a report .c.gcov.

-fprofile-filter-files=[regexes]

Define a list of regexes separated by a semi-colon. If a file name matches any of the regexes then the file is instrumented.

$ clang --coverage -fprofile-filter-files=".*\.c$" foo.c

For example, this will only instrument files finishing with .c, skipping .h files.

-fprofile-exclude-files=[regexes]

Define a list of regexes separated by a semi-colon. If a file name doesn’t match all the regexes then the file is instrumented.

$ clang --coverage -fprofile-exclude-files="^/usr/include/.*$" foo.c

For example, this will instrument all the files except the ones in /usr/include.

If both options are used then a file is instrumented if its name matches any of the regexes from -fprofile-filter-list and doesn’t match all the regexes from -fprofile-exclude-list.

$ clang --coverage -fprofile-exclude-files="^/usr/include/.*$" \
        -fprofile-filter-files="^/usr/.*$"

In that case /usr/foo/oof.h is instrumented since it matches the filter regex and doesn’t match the exclude regex, but /usr/include/foo.h doesn’t since it matches the exclude regex.

Controlling Debug Information

Controlling Size of Debug Information

Debug info kind generated by Clang can be set by one of the flags listed below. If multiple flags are present, the last one is used.

-g0

Don’t generate any debug info (default).

-gline-tables-only

Generate line number tables only.

This kind of debug info allows to obtain stack traces with function names, file names and line numbers (by such tools as gdb or addr2line). It doesn’t contain any other data (e.g. description of local variables or function parameters).

-fstandalone-debug

Clang supports a number of optimizations to reduce the size of debug information in the binary. They work based on the assumption that the debug type information can be spread out over multiple compilation units. Specifically, the optimizations are:

  • will not emit type definitions for types that are not needed by a module and could be replaced with a forward declaration.

  • will only emit type info for a dynamic C++ class in the module that contains the vtable for the class.

  • will only emit type info for a C++ class (non-trivial, non-aggregate) in the modules that contain a definition for one of its constructors.

  • will only emit type definitions for types that are the subject of explicit template instantiation declarations in the presence of an explicit instantiation definition for the type.

The -fstandalone-debug option turns off these optimizations. This is useful when working with 3rd-party libraries that don’t come with debug information. Note that Clang will never emit type information for types that are not referenced at all by the program.

-fno-standalone-debug

On Darwin -fstandalone-debug is enabled by default. The -fno-standalone-debug option can be used to get to turn on the vtable-based optimization described above.

-g

Generate complete debug info.

-feliminate-unused-debug-types

By default, Clang does not emit type information for types that are defined but not used in a program. To retain the debug info for these unused types, the negation -fno-eliminate-unused-debug-types can be used. This can be particulary useful on Windows, when using NATVIS files that can reference const symbols that would otherwise be stripped, even in full debug or standalone debug modes.

Controlling Macro Debug Info Generation

Debug info for C preprocessor macros increases the size of debug information in the binary. Macro debug info generated by Clang can be controlled by the flags listed below.

-fdebug-macro

Generate debug info for preprocessor macros. This flag is discarded when -g0 is enabled.

-fno-debug-macro

Do not generate debug info for preprocessor macros (default).

Controlling Debugger “Tuning”

While Clang generally emits standard DWARF debug info (http://dwarfstd.org), different debuggers may know how to take advantage of different specific DWARF features. You can “tune” the debug info for one of several different debuggers.

-ggdb, -glldb, -gsce, -gdbx

Tune the debug info for the gdb, lldb, Sony PlayStation® debugger, or dbx, respectively. Each of these options implies -g. (Therefore, if you want both -gline-tables-only and debugger tuning, the tuning option must come first.)

Controlling LLVM IR Output

Controlling Value Names in LLVM IR

Emitting value names in LLVM IR increases the size and verbosity of the IR. By default, value names are only emitted in assertion-enabled builds of Clang. However, when reading IR it can be useful to re-enable the emission of value names to improve readability.

-fdiscard-value-names

Discard value names when generating LLVM IR.

-fno-discard-value-names

Do not discard value names when generating LLVM IR. This option can be used to re-enable names for release builds of Clang.

Comment Parsing Options

Clang parses Doxygen and non-Doxygen style documentation comments and attaches them to the appropriate declaration nodes. By default, it only parses Doxygen-style comments and ignores ordinary comments starting with // and /*.

-Wdocumentation

Emit warnings about use of documentation comments. This warning group is off by default.

This includes checking that \param commands name parameters that actually present in the function signature, checking that \returns is used only on functions that actually return a value etc.

-Wno-documentation-unknown-command

Don’t warn when encountering an unknown Doxygen command.

-fparse-all-comments

Parse all comments as documentation comments (including ordinary comments starting with // and /*).

-fcomment-block-commands=[commands]

Define custom documentation commands as block commands. This allows Clang to construct the correct AST for these custom commands, and silences warnings about unknown commands. Several commands must be separated by a comma without trailing space; e.g. -fcomment-block-commands=foo,bar defines custom commands \foo and \bar.

It is also possible to use -fcomment-block-commands several times; e.g. -fcomment-block-commands=foo -fcomment-block-commands=bar does the same as above.

C Language Features

The support for standard C in clang is feature-complete except for the C99 floating-point pragmas.

Extensions supported by clang

See Clang Language Extensions.

Differences between various standard modes

clang supports the -std option, which changes what language mode clang uses. The supported modes for C are c89, gnu89, c94, c99, gnu99, c11, gnu11, c17, gnu17, c23, gnu23, c2y, gnu2y, and various aliases for those modes. If no -std option is specified, clang defaults to gnu17 mode. Many C99 and C11 features are supported in earlier modes as a conforming extension, with a warning. Use -pedantic-errors to request an error if a feature from a later standard revision is used in an earlier mode.

Differences between all c* and gnu* modes:

  • c* modes define “__STRICT_ANSI__”.

  • Target-specific defines not prefixed by underscores, like linux, are defined in gnu* modes.

  • Trigraphs default to being off in gnu* modes; they can be enabled by the -trigraphs option.

  • The parser recognizes asm and typeof as keywords in gnu* modes; the variants __asm__ and __typeof__ are recognized in all modes.

  • The parser recognizes inline as a keyword in gnu* mode, in addition to recognizing it in the *99 and later modes for which it is part of the ISO C standard. The variant __inline__ is recognized in all modes.

  • The Apple “blocks” extension is recognized by default in gnu* modes on some platforms; it can be enabled in any mode with the -fblocks option.

Differences between *89 and *94 modes:

  • Digraphs are not recognized in c89 mode.

Differences between *94 and *99 modes:

  • The *99 modes default to implementing inline / __inline__ as specified in C99, while the *89 modes implement the GNU version. This can be overridden for individual functions with the __gnu_inline__ attribute.

  • The scope of names defined inside a for, if, switch, while, or do statement is different. (example: if ((struct x {int x;}*)0) {}.)

  • __STDC_VERSION__ is not defined in *89 modes.

  • inline is not recognized as a keyword in c89 mode.

  • restrict is not recognized as a keyword in *89 modes.

  • Commas are allowed in integer constant expressions in *99 modes.

  • Arrays which are not lvalues are not implicitly promoted to pointers in *89 modes.

  • Some warnings are different.

Differences between *99 and *11 modes:

  • Warnings for use of C11 features are disabled.

  • __STDC_VERSION__ is defined to 201112L rather than 199901L.

Differences between *11 and *17 modes:

  • __STDC_VERSION__ is defined to 201710L rather than 201112L.

Differences between *17 and *23 modes:

  • __STDC_VERSION__ is defined to 202311L rather than 201710L.

  • nullptr and nullptr_t are supported, only in *23 mode.

  • ATOMIC_VAR_INIT is removed from *23 mode.

  • bool, true, false, alignas, alignof, static_assert, and thread_local are now first-class keywords, only in *23 mode.

  • typeof and typeof_unqual are supported, only *23 mode.

  • Bit-precise integers (_BitInt(N)) are supported by default in *23 mode, and as an extension in *17 and earlier modes.

  • [[]] attributes are supported by default in *23 mode, and as an extension in *17 and earlier modes.

Differences between *23 and *2y modes:

  • __STDC_VERSION__ is defined to 202400L rather than 202311L.

GCC extensions not implemented yet

clang tries to be compatible with gcc as much as possible, but some gcc extensions are not implemented yet:

  • clang does not support decimal floating point types (_Decimal32 and friends) yet.

  • clang does not support nested functions; this is a complex feature which is infrequently used, so it is unlikely to be implemented anytime soon. In C++11 it can be emulated by assigning lambda functions to local variables, e.g:

    auto const local_function = [&](int parameter) {
  // Do something
};
...
local_function(1);

  • clang only supports global register variables when the register specified is non-allocatable (e.g. the stack pointer). Support for general global register variables is unlikely to be implemented soon because it requires additional LLVM backend support.

  • clang does not support static initialization of flexible array members. This appears to be a rarely used extension, but could be implemented pending user demand.

  • clang does not support __builtin_va_arg_pack/__builtin_va_arg_pack_len. This is used rarely, but in some potentially interesting places, like the glibc headers, so it may be implemented pending user demand. Note that because clang pretends to be like GCC 4.2, and this extension was introduced in 4.3, the glibc headers will not try to use this extension with clang at the moment.

  • clang does not support the gcc extension for forward-declaring function parameters; this has not shown up in any real-world code yet, though, so it might never be implemented.

This is not a complete list; if you find an unsupported extension missing from this list, please send an e-mail to cfe-dev. This list currently excludes C++; see C++ Language Features. Also, this list does not include bugs in mostly-implemented features; please see the bug tracker for known existing bugs (FIXME: Is there a section for bug-reporting guidelines somewhere?).

Intentionally unsupported GCC extensions

  • clang does not support the gcc extension that allows variable-length arrays in structures. This is for a few reasons: one, it is tricky to implement, two, the extension is completely undocumented, and three, the extension appears to be rarely used. Note that clang does support flexible array members (arrays with a zero or unspecified size at the end of a structure).

  • GCC accepts many expression forms that are not valid integer constant expressions in bit-field widths, enumerator constants, case labels, and in array bounds at global scope. Clang also accepts additional expression forms in these contexts, but constructs that GCC accepts due to simplifications GCC performs while parsing, such as x - x (where x is a variable) will likely never be accepted by Clang.

  • clang does not support __builtin_apply and friends; this extension is extremely obscure and difficult to implement reliably.

Microsoft extensions

clang has support for many extensions from Microsoft Visual C++. To enable these extensions, use the -fms-extensions command-line option. This is the default for Windows targets. Clang does not implement every pragma or declspec provided by MSVC, but the popular ones, such as __declspec(dllexport) and #pragma comment(lib) are well supported.

clang has a -fms-compatibility flag that makes clang accept enough invalid C++ to be able to parse most Microsoft headers. For example, it allows unqualified lookup of dependent base class members, which is a common compatibility issue with clang. This flag is enabled by default for Windows targets.

-fdelayed-template-parsing lets clang delay parsing of function template definitions until the end of a translation unit. This flag is enabled by default for Windows targets.

For compatibility with existing code that compiles with MSVC, clang defines the _MSC_VER and _MSC_FULL_VER macros. When on Windows, these default to either the same value as the currently installed version of cl.exe, or 1933 and 193300000 (respectively). The -fms-compatibility-version= flag overrides these values. It accepts a dotted version tuple, such as 19.00.23506. Changing the MSVC compatibility version makes clang behave more like that version of MSVC. For example, -fms-compatibility-version=19 will enable C++14 features and define char16_t and char32_t as builtin types.

C++ Language Features

clang fully implements all of standard C++98 except for exported templates (which were removed in C++11), all of standard C++11, C++14, and C++17, and most of C++20.

See the C++ support in Clang page for detailed information on C++ feature support across Clang versions.

Controlling implementation limits

-fbracket-depth=N

Sets the limit for nested parentheses, brackets, and braces to N. The default is 256.

-fconstexpr-depth=N

Sets the limit for constexpr function invocations to N. The default is 512.

-fconstexpr-steps=N

Sets the limit for the number of full-expressions evaluated in a single constant expression evaluation. This also controls the maximum size of array and dynamic array allocation that can be constant evaluated. The default is 1048576.

-ftemplate-depth=N

Sets the limit for recursively nested template instantiations to N. The default is 1024.

-foperator-arrow-depth=N

Sets the limit for iterative calls to ‘operator->’ functions to N. The default is 256.

Objective-C Language Features

Objective-C++ Language Features

OpenMP Features

Clang supports all OpenMP 4.5 directives and clauses. See OpenMP Support for additional details.

Use -fopenmp to enable OpenMP. Support for OpenMP can be disabled with -fno-openmp.

Use -fopenmp-simd to enable OpenMP simd features only, without linking the runtime library; for combined constructs (e.g. #pragma omp parallel for simd) the non-simd directives and clauses will be ignored. This can be disabled with -fno-openmp-simd.

Controlling implementation limits

-fopenmp-use-tls

Controls code generation for OpenMP threadprivate variables. In presence of this option all threadprivate variables are generated the same way as thread local variables, using TLS support. If -fno-openmp-use-tls is provided or target does not support TLS, code generation for threadprivate variables relies on OpenMP runtime library.

OpenCL Features

Clang can be used to compile OpenCL kernels for execution on a device (e.g. GPU). It is possible to compile the kernel into a binary (e.g. for AMDGPU) that can be uploaded to run directly on a device (e.g. using clCreateProgramWithBinary) or into generic bitcode files loadable into other toolchains.

Compiling to a binary using the default target from the installation can be done as follows:

$ echo "kernel void k(){}" > test.cl
$ clang test.cl

Compiling for a specific target can be done by specifying the triple corresponding to the target, for example:

$ clang --target=nvptx64-unknown-unknown test.cl
$ clang --target=amdgcn-amd-amdhsa -mcpu=gfx900 test.cl

Compiling to bitcode can be done as follows:

$ clang -c -emit-llvm test.cl

This will produce a file test.bc that can be used in vendor toolchains to perform machine code generation.

Note that if compiled to bitcode for generic targets such as SPIR/SPIR-V, portable IR is produced that can be used with various vendor tools as well as open source tools such as SPIRV-LLVM Translator to produce SPIR-V binary. More details are provided in the offline compilation from OpenCL kernel sources into SPIR-V using open source tools. From clang 14 onwards SPIR-V can be generated directly as detailed in the SPIR-V support section.

Clang currently supports OpenCL C language standards up to v2.0. Clang mainly supports full profile. There is only very limited support of the embedded profile. From clang 9 a C++ mode is available for OpenCL (see C++ for OpenCL).

OpenCL v3.0 support is complete but it remains in experimental state, see more details about the experimental features and limitations in OpenCL Support page.

OpenCL Specific Options

Most of the OpenCL build options from the specification v2.0 section 5.8.4 are available.

Examples:

$ clang -cl-std=CL2.0 -cl-single-precision-constant test.cl

Many flags used for the compilation for C sources can also be passed while compiling for OpenCL, examples: -c, -O<1-4|s>, -o, -emit-llvm, etc.

Some extra options are available to support special OpenCL features.

-cl-no-stdinc

Allows to disable all extra types and functions that are not native to the compiler. This might reduce the compilation speed marginally but many declarations from the OpenCL standard will not be accessible. For example, the following will fail to compile.

$ echo "bool is_wg_uniform(int i){return get_enqueued_local_size(i)==get_local_size(i);}" > test.cl
$ clang -cl-std=CL2.0 -cl-no-stdinc test.cl
error: use of undeclared identifier 'get_enqueued_local_size'
error: use of undeclared identifier 'get_local_size'

More information about the standard types and functions is provided in the section on the OpenCL Header.

-cl-ext

Enables/Disables support of OpenCL extensions and optional features. All OpenCL targets set a list of extensions that they support. Clang allows to amend this using the -cl-ext flag with a comma-separated list of extensions prefixed with '+' or '-'. The syntax: -cl-ext=<(['-'|'+']<extension>[,])+>, where extensions can be either one of the OpenCL published extensions or any vendor extension. Alternatively, 'all' can be used to enable or disable all known extensions.

Example disabling double support for the 64-bit SPIR-V target:

$ clang -c --target=spirv64 -cl-ext=-cl_khr_fp64 test.cl

Enabling all extensions except double support in R600 AMD GPU can be done using:

$ clang --target=r600 -cl-ext=-all,+cl_khr_fp16 test.cl

Note that some generic targets e.g. SPIR/SPIR-V enable all extensions/features in clang by default.

OpenCL Targets

OpenCL targets are derived from the regular Clang target classes. The OpenCL specific parts of the target representation provide address space mapping as well as a set of supported extensions.

Specific Targets

There is a set of concrete HW architectures that OpenCL can be compiled for.

  • For AMD target:

    $ clang --target=amdgcn-amd-amdhsa -mcpu=gfx900 test.cl

  • For Nvidia architectures:

    $ clang --target=nvptx64-unknown-unknown test.cl

Generic Targets

  • A SPIR-V binary can be produced for 32 or 64 bit targets.

    $ clang --target=spirv32 -c test.cl
$ clang --target=spirv64 -c test.cl

    More details can be found in the SPIR-V support section.

  • SPIR is available as a generic target to allow portable bitcode to be produced that can be used across GPU toolchains. The implementation follows the SPIR specification. There are two flavors available for 32 and 64 bits.

    $ clang --target=spir test.cl -emit-llvm -c
$ clang --target=spir64 test.cl -emit-llvm -c

    Clang will generate SPIR v1.2 compatible IR for OpenCL versions up to 2.0 and SPIR v2.0 for OpenCL v2.0 or C++ for OpenCL.

  • x86 is used by some implementations that are x86 compatible and currently remains for backwards compatibility (with older implementations prior to SPIR target support). For “non-SPMD” targets which cannot spawn multiple work-items on the fly using hardware, which covers practically all non-GPU devices such as CPUs and DSPs, additional processing is needed for the kernels to support multiple work-item execution. For this, a 3rd party toolchain, such as for example POCL, can be used.

    This target does not support multiple memory segments and, therefore, the fake address space map can be added using the -ffake-address-space-map flag.

    All known OpenCL extensions and features are set to supported in the generic targets, however -cl-ext flag can be used to toggle individual extensions and features.

OpenCL Header

By default Clang will include standard headers and therefore most of OpenCL builtin functions and types are available during compilation. The default declarations of non-native compiler types and functions can be disabled by using flag -cl-no-stdinc.

The following example demonstrates that OpenCL kernel sources with various standard builtin functions can be compiled without the need for an explicit includes or compiler flags.

$ echo "bool is_wg_uniform(int i){return get_enqueued_local_size(i)==get_local_size(i);}" > test.cl
$ clang -cl-std=CL2.0 test.cl

More information about the default headers is provided in OpenCL Support.

OpenCL Extensions

Most of the cl_khr_* extensions to OpenCL C from the official OpenCL registry are available and configured per target depending on the support available in the specific architecture.

It is possible to alter the default extensions setting per target using -cl-ext flag. (See flags description for more details).

Vendor extensions can be added flexibly by declaring the list of types and functions associated with each extensions enclosed within the following compiler pragma directives:

#pragma OPENCL EXTENSION the_new_extension_name : begin
// declare types and functions associated with the extension here
#pragma OPENCL EXTENSION the_new_extension_name : end

For example, parsing the following code adds my_t type and my_func function to the custom my_ext extension.

#pragma OPENCL EXTENSION my_ext : begin
typedef struct{
  int a;
}my_t;
void my_func(my_t);
#pragma OPENCL EXTENSION my_ext : end

There is no conflict resolution for identifier clashes among extensions. It is therefore recommended that the identifiers are prefixed with a double underscore to avoid clashing with user space identifiers. Vendor extension should use reserved identifier prefix e.g. amd, arm, intel.

Clang also supports language extensions documented in The OpenCL C Language Extensions Documentation.

OpenCL-Specific Attributes

OpenCL support in Clang contains a set of attribute taken directly from the specification as well as additional attributes.

See also Attributes in Clang.

nosvm

Clang supports this attribute to comply to OpenCL v2.0 conformance, but it does not have any effect on the IR. For more details reffer to the specification section 6.7.2

opencl_unroll_hint

The implementation of this feature mirrors the unroll hint for C. More details on the syntax can be found in the specification section 6.11.5

convergent

To make sure no invalid optimizations occur for single program multiple data (SPMD) / single instruction multiple thread (SIMT) Clang provides attributes that can be used for special functions that have cross work item semantics. An example is the subgroup operations such as intel_sub_group_shuffle

// Define custom my_sub_group_shuffle(data, c)
// that makes use of intel_sub_group_shuffle
r1 = ...
if (r0) r1 = computeA();
// Shuffle data from r1 into r3
// of threads id r2.
r3 = my_sub_group_shuffle(r1, r2);
if (r0) r3 = computeB();

with non-SPMD semantics this is optimized to the following equivalent code:

r1 = ...
if (!r0)
  // Incorrect functionality! The data in r1
  // have not been computed by all threads yet.
  r3 = my_sub_group_shuffle(r1, r2);
else {
  r1 = computeA();
  r3 = my_sub_group_shuffle(r1, r2);
  r3 = computeB();
}

Declaring the function my_sub_group_shuffle with the convergent attribute would prevent this:

my_sub_group_shuffle() __attribute__((convergent));

Using convergent guarantees correct execution by keeping CFG equivalence wrt operations marked as convergent. CFG G´ is equivalent to G wrt node Ni : iff ∀ Nj (i≠j) domination and post-domination relations with respect to Ni remain the same in both G and G´.

noduplicate

noduplicate is more restrictive with respect to optimizations than convergent because a convergent function only preserves CFG equivalence. This allows some optimizations to happen as long as the control flow remains unmodified.

for (int i=0; i<4; i++)
  my_sub_group_shuffle()

can be modified to:

my_sub_group_shuffle();
my_sub_group_shuffle();
my_sub_group_shuffle();
my_sub_group_shuffle();

while using noduplicate would disallow this. Also noduplicate doesn’t have the same safe semantics of CFG as convergent and can cause changes in CFG that modify semantics of the original program.

noduplicate is kept for backwards compatibility only and it considered to be deprecated for future uses.

C++ for OpenCL

Starting from clang 9 kernel code can contain C++17 features: classes, templates, function overloading, type deduction, etc. Please note that this is not an implementation of OpenCL C++ and there is no plan to support it in clang in any new releases in the near future.

Clang currently supports C++ for OpenCL 1.0 and 2021. For detailed information about this language refer to the C++ for OpenCL Programming Language Documentation available in the latest build or in the official release.

To enable the C++ for OpenCL mode, pass one of following command line options when compiling .clcpp file:

  • C++ for OpenCL 1.0: -cl-std=clc++, -cl-std=CLC++, -cl-std=clc++1.0, -cl-std=CLC++1.0, -std=clc++, -std=CLC++, -std=clc++1.0 or -std=CLC++1.0.

  • C++ for OpenCL 2021: -cl-std=clc++2021, -cl-std=CLC++2021, -std=clc++2021, -std=CLC++2021.

Example of use:
template<class T> T add( T x, T y )
{
  return x + y;
}

__kernel void test( __global float* a, __global float* b)
{
  auto index = get_global_id(0);
  a[index] = add(b[index], b[index+1]);
}

clang -cl-std=clc++1.0 test.clcpp
clang -cl-std=clc++ -c --target=spirv64 test.cl

By default, files with .clcpp extension are compiled with the C++ for OpenCL 1.0 mode.

clang test.clcpp

For backward compatibility files with .cl extensions can also be compiled in C++ for OpenCL mode but the desirable language mode must be activated with a flag.

clang -cl-std=clc++ test.cl

Support of C++ for OpenCL 2021 is currently in experimental phase, refer to OpenCL Support for more details.

C++ for OpenCL kernel sources can also be compiled online in drivers supporting cl_ext_cxx_for_opencl extension.

Constructing and destroying global objects

Global objects with non-trivial constructors require the constructors to be run before the first kernel using the global objects is executed. Similarly global objects with non-trivial destructors require destructor invocation just after the last kernel using the program objects is executed. In OpenCL versions earlier than v2.2 there is no support for invoking global constructors. However, an easy workaround is to manually enqueue the constructor initialization kernel that has the following name scheme _GLOBAL__sub_I_<compiled file name>. This kernel is only present if there are global objects with non-trivial constructors present in the compiled binary. One way to check this is by passing CL_PROGRAM_KERNEL_NAMES to clGetProgramInfo (OpenCL v2.0 s5.8.7) and then checking whether any kernel name matches the naming scheme of global constructor initialization kernel above.

Note that if multiple files are compiled and linked into libraries, multiple kernels that initialize global objects for multiple modules would have to be invoked.

Applications are currently required to run initialization of global objects manually before running any kernels in which the objects are used.

clang -cl-std=clc++ test.cl

If there are any global objects to be initialized, the final binary will contain the _GLOBAL__sub_I_test.cl kernel to be enqueued.

Note that the manual workaround only applies to objects declared at the program scope. There is no manual workaround for the construction of static objects with non-trivial constructors inside functions.

Global destructors can not be invoked manually in the OpenCL v2.0 drivers. However, all memory used for program scope objects should be released on clReleaseProgram.

Libraries

Limited experimental support of C++ standard libraries for OpenCL is described in OpenCL Support page.

Target-Specific Features and Limitations

CPU Architectures Features and Limitations

X86

The support for X86 (both 32-bit and 64-bit) is considered stable on Darwin (macOS), Linux, FreeBSD, and Dragonfly BSD: it has been tested to correctly compile many large C, C++, Objective-C, and Objective-C++ codebases.

On x86_64-mingw32, passing i128(by value) is incompatible with the Microsoft x64 calling convention. You might need to tweak WinX86_64ABIInfo::classify() in lib/CodeGen/Targets/X86.cpp.

For the X86 target, clang supports the -m16 command line argument which enables 16-bit code output. This is broadly similar to using asm(".code16gcc") with the GNU toolchain. The generated code and the ABI remains 32-bit but the assembler emits instructions appropriate for a CPU running in 16-bit mode, with address-size and operand-size prefixes to enable 32-bit addressing and operations.

Several micro-architecture levels as specified by the x86-64 psABI are defined. They are cumulative in the sense that features from previous levels are implicitly included in later levels.

  • -march=x86-64: CMOV, CMPXCHG8B, FPU, FXSR, MMX, FXSR, SCE, SSE, SSE2

  • -march=x86-64-v2: (close to Nehalem) CMPXCHG16B, LAHF-SAHF, POPCNT, SSE3, SSE4.1, SSE4.2, SSSE3

  • -march=x86-64-v3: (close to Haswell) AVX, AVX2, BMI1, BMI2, F16C, FMA, LZCNT, MOVBE, XSAVE

  • -march=x86-64-v4: AVX512F, AVX512BW, AVX512CD, AVX512DQ, AVX512VL

Intel AVX10 ISA is a major new vector ISA incorporating the modern vectorization aspects of Intel AVX-512. This ISA will be supported on all future Intel processors. Users are supposed to use the new options -mavx10.N and -mavx10.N-512 on these processors and should not use traditional AVX512 options anymore.

The N in -mavx10.N represents a continuous integer number starting from 1. -mavx10.N is an alias of -mavx10.N-256, which means to enable all instructions within AVX10 version N at a maximum vector length of 256 bits. -mavx10.N-512 enables all instructions at a maximum vector length of 512 bits, which is a superset of instructions -mavx10.N enabled.

Current binaries built with AVX512 features can run on Intel AVX10/512 capable processors without re-compile, but cannot run on AVX10/256 capable processors. Users need to re-compile their code with -mavx10.N, and maybe update some code that calling to 512-bit X86 specific intrinsics and passing or returning 512-bit vector types in function call, if they want to run on AVX10/256 capable processors. Binaries built with -mavx10.N can run on both AVX10/256 and AVX10/512 capable processors.

Users can add a -mno-evex512 in the command line with AVX512 options if they want to run the binary on both legacy AVX512 and new AVX10/256 capable processors. The option has the same constraints as -mavx10.N, i.e., cannot call to 512-bit X86 specific intrinsics and pass or return 512-bit vector types in function call.

Users should avoid using AVX512 features in function target attributes when developing code for AVX10. If they have to do so, they need to add an explicit evex512 or no-evex512 together with AVX512 features for 512-bit or non-512-bit functions respectively to avoid unexpected code generation. Both command line option and target attribute of EVEX512 feature can only be used with AVX512. They don’t affect vector size of AVX10.

User should not mix the use AVX10 and AVX512 options together at any time, because the option combinations are conflicting sometimes. For example, a combination of -mavx512f -mavx10.1-256 doesn’t show a clear intention to compiler, since instructions in AVX512F and AVX10.1/256 intersect but do not overlap. In this case, compiler will emit warning for it, but the behavior is determined. It will generate the same code as option -mavx10.1-512. A similar case is -mavx512f -mavx10.2-256, which equals to -mavx10.1-512 -mavx10.2-256, because avx10.2-256 implies avx10.1-256 and -mavx512f -mavx10.1-256 equals to -mavx10.1-512.

There are some new macros introduced with AVX10 support. -mavx10.1-256 will enable __AVX10_1__ and __EVEX256__, while -mavx10.1-512 enables __AVX10_1__, __EVEX256__, __EVEX512__ and __AVX10_1_512__. Besides, both -mavx10.1-256 and -mavx10.1-512 will enable all AVX512 feature specific macros. A AVX512 feature will enable both __EVEX256__, __EVEX512__ and its own macro. So __EVEX512__ can be used to guard code that can run on both legacy AVX512 and AVX10/512 capable processors but cannot run on AVX10/256, while a AVX512 macro like __AVX512F__ cannot tell the difference among the three options. Users need to check additional macros __AVX10_1__ and __EVEX512__ if they want to make distinction.

ARM

The support for ARM (specifically ARMv6 and ARMv7) is considered stable on Darwin (iOS): it has been tested to correctly compile many large C, C++, Objective-C, and Objective-C++ codebases. Clang only supports a limited number of ARM architectures. It does not yet fully support ARMv5, for example.

PowerPC

The support for PowerPC (especially PowerPC64) is considered stable on Linux and FreeBSD: it has been tested to correctly compile many large C and C++ codebases. PowerPC (32bit) is still missing certain features (e.g. PIC code on ELF platforms).

Other platforms

clang currently contains some support for other architectures (e.g. Sparc); however, significant pieces of code generation are still missing, and they haven’t undergone significant testing.

clang contains limited support for the MSP430 embedded processor, but both the clang support and the LLVM backend support are highly experimental.

Other platforms are completely unsupported at the moment. Adding the minimal support needed for parsing and semantic analysis on a new platform is quite easy; see lib/Basic/Targets.cpp in the clang source tree. This level of support is also sufficient for conversion to LLVM IR for simple programs. Proper support for conversion to LLVM IR requires adding code to lib/CodeGen/CGCall.cpp at the moment; this is likely to change soon, though. Generating assembly requires a suitable LLVM backend.

Operating System Features and Limitations

Windows

Clang has experimental support for targeting “Cygming” (Cygwin / MinGW) platforms.

See also Microsoft Extensions.

Cygwin

Clang works on Cygwin-1.7.

MinGW32

Clang works on some mingw32 distributions. Clang assumes directories as below;

  • C:/mingw/include

  • C:/mingw/lib

  • C:/mingw/lib/gcc/mingw32/4.[3-5].0/include/c++

On MSYS, a few tests might fail.

MinGW-w64

For 32-bit (i686-w64-mingw32), and 64-bit (x86_64-w64-mingw32), Clang assumes as below;

  • GCC versions 4.5.0 to 4.5.3, 4.6.0 to 4.6.2, or 4.7.0 (for the C++ header search path)

  • some_directory/bin/gcc.exe

  • some_directory/bin/clang.exe

  • some_directory/bin/clang++.exe

  • some_directory/bin/../include/c++/GCC_version

  • some_directory/bin/../include/c++/GCC_version/x86_64-w64-mingw32

  • some_directory/bin/../include/c++/GCC_version/i686-w64-mingw32

  • some_directory/bin/../include/c++/GCC_version/backward

  • some_directory/bin/../x86_64-w64-mingw32/include

  • some_directory/bin/../i686-w64-mingw32/include

  • some_directory/bin/../include

This directory layout is standard for any toolchain you will find on the official MinGW-w64 website.

Clang expects the GCC executable “gcc.exe” compiled for i686-w64-mingw32 (or x86_64-w64-mingw32) to be present on PATH.

Some tests might fail on x86_64-w64-mingw32.

AIX

TOC Data Transformation

TOC data transformation is off by default (-mno-tocdata). When -mtocdata is specified, the TOC data transformation will be applied to all suitable variables with static storage duration, including static data members of classes and block-scope static variables (if not marked as exceptions, see further below).

Suitable variables must:

  • have complete types

  • be independently generated (i.e., not placed in a pool)

  • be at most as large as a pointer

  • not be aligned more strictly than a pointer

  • not be structs containing flexible array members

  • not have internal linkage

  • not have aliases

  • not have section attributes

  • not be thread local storage

The TOC data transformation results in the variable, not its address, being placed in the TOC. This eliminates the need to load the address of the variable from the TOC.

Note: If the TOC data transformation is applied to a variable whose definition is imported, the linker will generate fixup code for reading or writing to the variable.

When multiple toc-data options are used, the last option used has the affect. For example: -mno-tocdata=g5,g1 -mtocdata=g1,g2 -mno-tocdata=g2 -mtocdata=g3,g4 results in -mtocdata=g1,g3,g4

Names of variables not having external linkage will be ignored.

Options:

-mno-tocdata

This is the default behaviour. Only variables explicitly specified with -mtocdata= will have the TOC data transformation applied.

-mtocdata

Apply the TOC data transformation to all suitable variables with static storage duration (including static data members of classes and block-scope static variables) that are not explicitly specified with -mno-tocdata=.

-mno-tocdata=

Can be used in conjunction with -mtocdata to mark the comma-separated list of external linkage variables, specified using their mangled names, as exceptions to -mtocdata.

-mtocdata=

Apply the TOC data transformation to the comma-separated list of external linkage variables, specified using their mangled names, if they are suitable. Emit diagnostics for all unsuitable variables specified.

Default Visibility Export Mapping

The -mdefault-visibility-export-mapping= option can be used to control mapping of default visibility to an explicit shared object export (i.e. XCOFF exported visibility). Three values are provided for the option:

  • -mdefault-visibility-export-mapping=none: no additional export information is created for entities with default visibility.

  • -mdefault-visibility-export-mapping=explicit: mark entities for export if they have explicit (e.g. via an attribute) default visibility from the source, including RTTI.

  • -mdefault-visibility-export-mapping=all: set XCOFF exported visibility for all entities with default visibility from any source. This gives a export behavior similar to ELF platforms where all entities with default visibility are exported.

SPIR-V support

Clang supports generation of SPIR-V conformant to the OpenCL Environment Specification.

To generate SPIR-V binaries, Clang uses the external llvm-spirv tool from the SPIRV-LLVM-Translator repo.

Prior to the generation of SPIR-V binary with Clang, llvm-spirv should be built or installed. Please refer to the following instructions for more details. Clang will look for llvm-spirv-<LLVM-major-version> and llvm-spirv executables, in this order, in the PATH environment variable. Clang uses llvm-spirv with the widely adopted assembly syntax package.

The versioning of llvm-spirv is aligned with Clang major releases. The same applies to the main development branch. It is therefore important to ensure the llvm-spirv version is in alignment with the Clang version. For troubleshooting purposes llvm-spirv can be tested in isolation.

Example usage for OpenCL kernel compilation:

$ clang --target=spirv32 -c test.cl
$ clang --target=spirv64 -c test.cl

Both invocations of Clang will result in the generation of a SPIR-V binary file test.o for 32 bit and 64 bit respectively. This file can be imported by an OpenCL driver that support SPIR-V consumption or it can be compiled further by offline SPIR-V consumer tools.

Converting to SPIR-V produced with the optimization levels other than -O0 is currently available as an experimental feature and it is not guaranteed to work in all cases.

Clang also supports integrated generation of SPIR-V without use of llvm-spirv tool as an experimental feature when -fintegrated-objemitter flag is passed in the command line.

$ clang --target=spirv32 -fintegrated-objemitter -c test.cl

Note that only very basic functionality is supported at this point and therefore it is not suitable for arbitrary use cases. This feature is only enabled when clang build is configured with -DLLVM_EXPERIMENTAL_TARGETS_TO_BUILD=SPIRV option.

Linking is done using spirv-link from the SPIRV-Tools project. Similar to other external linkers, Clang will expect spirv-link to be installed separately and to be present in the PATH environment variable. Please refer to the build and installation instructions.

$ clang --target=spirv64 test1.cl test2.cl

More information about the SPIR-V target settings and supported versions of SPIR-V format can be found in the SPIR-V target guide.

clang-cl

clang-cl is an alternative command-line interface to Clang, designed for compatibility with the Visual C++ compiler, cl.exe.

To enable clang-cl to find system headers, libraries, and the linker when run from the command-line, it should be executed inside a Visual Studio Native Tools Command Prompt or a regular Command Prompt where the environment has been set up using e.g. vcvarsall.bat.

clang-cl can also be used from inside Visual Studio by selecting the LLVM Platform Toolset. The toolset is not part of the installer, but may be installed separately from the Visual Studio Marketplace. To use the toolset, select a project in Solution Explorer, open its Property Page (Alt+F7), and in the “General” section of “Configuration Properties” change “Platform Toolset” to LLVM. Doing so enables an additional Property Page for selecting the clang-cl executable to use for builds.

To use the toolset with MSBuild directly, invoke it with e.g. /p:PlatformToolset=LLVM. This allows trying out the clang-cl toolchain without modifying your project files.

It’s also possible to point MSBuild at clang-cl without changing toolset by passing /p:CLToolPath=c:\llvm\bin /p:CLToolExe=clang-cl.exe.

When using CMake and the Visual Studio generators, the toolset can be set with the -T flag:

cmake -G"Visual Studio 16 2019" -T LLVM ..

When using CMake with the Ninja generator, set the CMAKE_C_COMPILER and CMAKE_CXX_COMPILER variables to clang-cl:

cmake -GNinja -DCMAKE_C_COMPILER="c:/Program Files (x86)/LLVM/bin/clang-cl.exe"
    -DCMAKE_CXX_COMPILER="c:/Program Files (x86)/LLVM/bin/clang-cl.exe" ..

Command-Line Options

To be compatible with cl.exe, clang-cl supports most of the same command-line options. Those options can start with either / or -. It also supports some of Clang’s core options, such as the -W options.

Options that are known to clang-cl, but not currently supported, are ignored with a warning. For example:

clang-cl.exe: warning: argument unused during compilation: '/AI'

To suppress warnings about unused arguments, use the -Qunused-arguments option.

Options that are not known to clang-cl will be ignored by default. Use the -Werror=unknown-argument option in order to treat them as errors. If these options are spelled with a leading /, they will be mistaken for a filename:

clang-cl.exe: error: no such file or directory: '/foobar'

Please file a bug for any valid cl.exe flags that clang-cl does not understand.

Execute clang-cl /? to see a list of supported options:

CL.EXE COMPATIBILITY OPTIONS:
  /?                      Display available options
  /arch:<value>           Set architecture for code generation
  /Brepro-                Emit an object file which cannot be reproduced over time
  /Brepro                 Emit an object file which can be reproduced over time
  /clang:<arg>            Pass <arg> to the clang driver
  /C                      Don't discard comments when preprocessing
  /c                      Compile only
  /d1PP                   Retain macro definitions in /E mode
  /d1reportAllClassLayout Dump record layout information
  /diagnostics:caret      Enable caret and column diagnostics (on by default)
  /diagnostics:classic    Disable column and caret diagnostics
  /diagnostics:column     Disable caret diagnostics but keep column info
  /D <macro[=value]>      Define macro
  /EH<value>              Exception handling model
  /EP                     Disable linemarker output and preprocess to stdout
  /execution-charset:<value>
                          Runtime encoding, supports only UTF-8
  /E                      Preprocess to stdout
  /FA                     Output assembly code file during compilation
  /Fa<file or directory>  Output assembly code to this file during compilation (with /FA)
  /Fe<file or directory>  Set output executable file or directory (ends in / or \)
  /FI <value>             Include file before parsing
  /Fi<file>               Set preprocess output file name (with /P)
  /Fo<file or directory>  Set output object file, or directory (ends in / or \) (with /c)
  /fp:except-
  /fp:except
  /fp:fast
  /fp:precise
  /fp:strict
  /Fp<filename>           Set pch filename (with /Yc and /Yu)
  /GA                     Assume thread-local variables are defined in the executable
  /Gd                     Set __cdecl as a default calling convention
  /GF-                    Disable string pooling
  /GF                     Enable string pooling (default)
  /GR-                    Disable emission of RTTI data
  /Gregcall               Set __regcall as a default calling convention
  /GR                     Enable emission of RTTI data
  /Gr                     Set __fastcall as a default calling convention
  /GS-                    Disable buffer security check
  /GS                     Enable buffer security check (default)
  /Gs                     Use stack probes (default)
  /Gs<value>              Set stack probe size (default 4096)
  /guard:<value>          Enable Control Flow Guard with /guard:cf,
                          or only the table with /guard:cf,nochecks.
                          Enable EH Continuation Guard with /guard:ehcont
  /Gv                     Set __vectorcall as a default calling convention
  /Gw-                    Don't put each data item in its own section
  /Gw                     Put each data item in its own section
  /GX-                    Disable exception handling
  /GX                     Enable exception handling
  /Gy-                    Don't put each function in its own section (default)
  /Gy                     Put each function in its own section
  /Gz                     Set __stdcall as a default calling convention
  /help                   Display available options
  /imsvc <dir>            Add directory to system include search path, as if part of %INCLUDE%
  /I <dir>                Add directory to include search path
  /J                      Make char type unsigned
  /LDd                    Create debug DLL
  /LD                     Create DLL
  /link <options>         Forward options to the linker
  /MDd                    Use DLL debug run-time
  /MD                     Use DLL run-time
  /MTd                    Use static debug run-time
  /MT                     Use static run-time
  /O0                     Disable optimization
  /O1                     Optimize for size  (same as /Og     /Os /Oy /Ob2 /GF /Gy)
  /O2                     Optimize for speed (same as /Og /Oi /Ot /Oy /Ob2 /GF /Gy)
  /Ob0                    Disable function inlining
  /Ob1                    Only inline functions which are (explicitly or implicitly) marked inline
  /Ob2                    Inline functions as deemed beneficial by the compiler
  /Ob3                    Same as /Ob2
  /Od                     Disable optimization
  /Og                     No effect
  /Oi-                    Disable use of builtin functions
  /Oi                     Enable use of builtin functions
  /Os                     Optimize for size (like clang -Os)
  /Ot                     Optimize for speed (like clang -O3)
  /Ox                     Deprecated (same as /Og /Oi /Ot /Oy /Ob2); use /O2 instead
  /Oy-                    Disable frame pointer omission (x86 only, default)
  /Oy                     Enable frame pointer omission (x86 only)
  /O<flags>               Set multiple /O flags at once; e.g. '/O2y-' for '/O2 /Oy-'
  /o <file or directory>  Set output file or directory (ends in / or \)
  /P                      Preprocess to file
  /Qvec-                  Disable the loop vectorization passes
  /Qvec                   Enable the loop vectorization passes
  /showFilenames-         Don't print the name of each compiled file (default)
  /showFilenames          Print the name of each compiled file
  /showIncludes           Print info about included files to stderr
  /source-charset:<value> Source encoding, supports only UTF-8
  /std:<value>            Language standard to compile for
  /TC                     Treat all source files as C
  /Tc <filename>          Specify a C source file
  /TP                     Treat all source files as C++
  /Tp <filename>          Specify a C++ source file
  /utf-8                  Set source and runtime encoding to UTF-8 (default)
  /U <macro>              Undefine macro
  /vd<value>              Control vtordisp placement
  /vmb                    Use a best-case representation method for member pointers
  /vmg                    Use a most-general representation for member pointers
  /vmm                    Set the default most-general representation to multiple inheritance
  /vms                    Set the default most-general representation to single inheritance
  /vmv                    Set the default most-general representation to virtual inheritance
  /volatile:iso           Volatile loads and stores have standard semantics
  /volatile:ms            Volatile loads and stores have acquire and release semantics
  /W0                     Disable all warnings
  /W1                     Enable -Wall
  /W2                     Enable -Wall
  /W3                     Enable -Wall
  /W4                     Enable -Wall and -Wextra
  /Wall                   Enable -Weverything
  /WX-                    Do not treat warnings as errors
  /WX                     Treat warnings as errors
  /w                      Disable all warnings
  /X                      Don't add %INCLUDE% to the include search path
  /Y-                     Disable precompiled headers, overrides /Yc and /Yu
  /Yc<filename>           Generate a pch file for all code up to and including <filename>
  /Yu<filename>           Load a pch file and use it instead of all code up to and including <filename>
  /Z7                     Enable CodeView debug information in object files
  /Zc:char8_t             Enable C++20 char8_t type
  /Zc:char8_t-            Disable C++20 char8_t type
  /Zc:dllexportInlines-   Don't dllexport/dllimport inline member functions of dllexport/import classes
  /Zc:dllexportInlines    dllexport/dllimport inline member functions of dllexport/import classes (default)
  /Zc:sizedDealloc-       Disable C++14 sized global deallocation functions
  /Zc:sizedDealloc        Enable C++14 sized global deallocation functions
  /Zc:strictStrings       Treat string literals as const
  /Zc:threadSafeInit-     Disable thread-safe initialization of static variables
  /Zc:threadSafeInit      Enable thread-safe initialization of static variables
  /Zc:trigraphs-          Disable trigraphs (default)
  /Zc:trigraphs           Enable trigraphs
  /Zc:twoPhase-           Disable two-phase name lookup in templates
  /Zc:twoPhase            Enable two-phase name lookup in templates
  /Zi                     Alias for /Z7. Does not produce PDBs.
  /Zl                     Don't mention any default libraries in the object file
  /Zp                     Set the default maximum struct packing alignment to 1
  /Zp<value>              Specify the default maximum struct packing alignment
  /Zs                     Run the preprocessor, parser and semantic analysis stages

OPTIONS:
  -###                    Print (but do not run) the commands to run for this compilation
  --analyze               Run the static analyzer
  -faddrsig               Emit an address-significance table
  -fansi-escape-codes     Use ANSI escape codes for diagnostics
  -fblocks                Enable the 'blocks' language feature
  -fcf-protection=<value> Instrument control-flow architecture protection. Options: return, branch, full, none.
  -fcf-protection         Enable cf-protection in 'full' mode
  -fcolor-diagnostics     Use colors in diagnostics
  -fcomplete-member-pointers
                          Require member pointer base types to be complete if they would be significant under the Microsoft ABI
  -fcoverage-mapping      Generate coverage mapping to enable code coverage analysis
  -fcrash-diagnostics-dir=<dir>
                          Put crash-report files in <dir>
  -fdebug-macro           Emit macro debug information
  -fdelayed-template-parsing
                          Parse templated function definitions at the end of the translation unit
  -fdiagnostics-absolute-paths
                          Print absolute paths in diagnostics
  -fdiagnostics-parseable-fixits
                          Print fix-its in machine parseable form
  -flto=<value>           Set LTO mode to either 'full' or 'thin'
  -flto                   Enable LTO in 'full' mode
  -fmerge-all-constants   Allow merging of constants
  -fmodule-file=<module_name>=<module-file>
                          Use the specified module file that provides the module <module_name>
  -fmodule-header=<header>
                          Build <header> as a C++20 header unit
  -fmodule-output=<path>
                          Save intermediate module file results when compiling a standard C++ module unit.
  -fms-compatibility-version=<value>
                          Dot-separated value representing the Microsoft compiler version
                          number to report in _MSC_VER (0 = don't define it; default is same value as installed cl.exe, or 1933)
  -fms-compatibility      Enable full Microsoft Visual C++ compatibility
  -fms-extensions         Accept some non-standard constructs supported by the Microsoft compiler
  -fmsc-version=<value>   Microsoft compiler version number to report in _MSC_VER
                          (0 = don't define it; default is same value as installed cl.exe, or 1933)
  -fno-addrsig            Don't emit an address-significance table
  -fno-builtin-<value>    Disable implicit builtin knowledge of a specific function
  -fno-builtin            Disable implicit builtin knowledge of functions
  -fno-complete-member-pointers
                          Do not require member pointer base types to be complete if they would be significant under the Microsoft ABI
  -fno-coverage-mapping   Disable code coverage analysis
  -fno-crash-diagnostics  Disable auto-generation of preprocessed source files and a script for reproduction during a clang crash
  -fno-debug-macro        Do not emit macro debug information
  -fno-delayed-template-parsing
                          Disable delayed template parsing
  -fno-sanitize-address-poison-custom-array-cookie
                          Disable poisoning array cookies when using custom operator new[] in AddressSanitizer
  -fno-sanitize-address-use-after-scope
                          Disable use-after-scope detection in AddressSanitizer
  -fno-sanitize-address-use-odr-indicator
                           Disable ODR indicator globals
  -fno-sanitize-ignorelist Don't use ignorelist file for sanitizers
  -fno-sanitize-cfi-cross-dso
                          Disable control flow integrity (CFI) checks for cross-DSO calls.
  -fno-sanitize-coverage=<value>
                          Disable specified features of coverage instrumentation for Sanitizers
  -fno-sanitize-memory-track-origins
                          Disable origins tracking in MemorySanitizer
  -fno-sanitize-memory-use-after-dtor
                          Disable use-after-destroy detection in MemorySanitizer
  -fno-sanitize-recover=<value>
                          Disable recovery for specified sanitizers
  -fno-sanitize-stats     Disable sanitizer statistics gathering.
  -fno-sanitize-thread-atomics
                          Disable atomic operations instrumentation in ThreadSanitizer
  -fno-sanitize-thread-func-entry-exit
                          Disable function entry/exit instrumentation in ThreadSanitizer
  -fno-sanitize-thread-memory-access
                          Disable memory access instrumentation in ThreadSanitizer
  -fno-sanitize-trap=<value>
                          Disable trapping for specified sanitizers
  -fno-standalone-debug   Limit debug information produced to reduce size of debug binary
  -fno-strict-aliasing    Disable optimizations based on strict aliasing rules (default)
  -fobjc-runtime=<value>  Specify the target Objective-C runtime kind and version
  -fprofile-exclude-files=<value>
                          Instrument only functions from files where names don't match all the regexes separated by a semi-colon
  -fprofile-filter-files=<value>
                          Instrument only functions from files where names match any regex separated by a semi-colon
  -fprofile-generate=<dirname>
                          Generate instrumented code to collect execution counts into a raw profile file in the directory specified by the argument. The filename uses default_%m.profraw pattern
                          (overridden by LLVM_PROFILE_FILE env var)
  -fprofile-generate
                          Generate instrumented code to collect execution counts into default_%m.profraw file
                          (overridden by '=' form of option or LLVM_PROFILE_FILE env var)
  -fprofile-instr-generate=<file_name_pattern>
                          Generate instrumented code to collect execution counts into the file whose name pattern is specified as the argument
                          (overridden by LLVM_PROFILE_FILE env var)
  -fprofile-instr-generate
                          Generate instrumented code to collect execution counts into default.profraw file
                          (overridden by '=' form of option or LLVM_PROFILE_FILE env var)
  -fprofile-instr-use=<value>
                          Use instrumentation data for coverage testing or profile-guided optimization
  -fprofile-use=<value>
                          Use instrumentation data for profile-guided optimization
  -fprofile-remapping-file=<file>
                          Use the remappings described in <file> to match the profile data against names in the program
  -fprofile-list=<file>
                          Filename defining the list of functions/files to instrument
  -fsanitize-address-field-padding=<value>
                          Level of field padding for AddressSanitizer
  -fsanitize-address-globals-dead-stripping
                          Enable linker dead stripping of globals in AddressSanitizer
  -fsanitize-address-poison-custom-array-cookie
                          Enable poisoning array cookies when using custom operator new[] in AddressSanitizer
  -fsanitize-address-use-after-return=<mode>
                          Select the mode of detecting stack use-after-return in AddressSanitizer: never | runtime (default) | always
  -fsanitize-address-use-after-scope
                          Enable use-after-scope detection in AddressSanitizer
  -fsanitize-address-use-odr-indicator
                          Enable ODR indicator globals to avoid false ODR violation reports in partially sanitized programs at the cost of an increase in binary size
  -fsanitize-ignorelist=<value>
                          Path to ignorelist file for sanitizers
  -fsanitize-cfi-cross-dso
                          Enable control flow integrity (CFI) checks for cross-DSO calls.
  -fsanitize-cfi-icall-generalize-pointers
                          Generalize pointers in CFI indirect call type signature checks
  -fsanitize-coverage=<value>
                          Specify the type of coverage instrumentation for Sanitizers
  -fsanitize-hwaddress-abi=<value>
                          Select the HWAddressSanitizer ABI to target (interceptor or platform, default interceptor)
  -fsanitize-memory-track-origins=<value>
                          Enable origins tracking in MemorySanitizer
  -fsanitize-memory-track-origins
                          Enable origins tracking in MemorySanitizer
  -fsanitize-memory-use-after-dtor
                          Enable use-after-destroy detection in MemorySanitizer
  -fsanitize-recover=<value>
                          Enable recovery for specified sanitizers
  -fsanitize-stats        Enable sanitizer statistics gathering.
  -fsanitize-thread-atomics
                          Enable atomic operations instrumentation in ThreadSanitizer (default)
  -fsanitize-thread-func-entry-exit
                          Enable function entry/exit instrumentation in ThreadSanitizer (default)
  -fsanitize-thread-memory-access
                          Enable memory access instrumentation in ThreadSanitizer (default)
  -fsanitize-trap=<value> Enable trapping for specified sanitizers
  -fsanitize-undefined-strip-path-components=<number>
                          Strip (or keep only, if negative) a given number of path components when emitting check metadata.
  -fsanitize=<check>      Turn on runtime checks for various forms of undefined or suspicious
                          behavior. See user manual for available checks
  -fsplit-lto-unit        Enables splitting of the LTO unit.
  -fstandalone-debug      Emit full debug info for all types used by the program
  -fstrict-aliasing       Enable optimizations based on strict aliasing rules
  -fsyntax-only           Run the preprocessor, parser and semantic analysis stages
  -fwhole-program-vtables Enables whole-program vtable optimization. Requires -flto
  -gcodeview-ghash        Emit type record hashes in a .debug$H section
  -gcodeview              Generate CodeView debug information
  -gline-directives-only  Emit debug line info directives only
  -gline-tables-only      Emit debug line number tables only
  -miamcu                 Use Intel MCU ABI
  -mllvm <value>          Additional arguments to forward to LLVM's option processing
  -nobuiltininc           Disable builtin #include directories
  -Qunused-arguments      Don't emit warning for unused driver arguments
  -R<remark>              Enable the specified remark
  --target=<value>        Generate code for the given target
  --version               Print version information
  -v                      Show commands to run and use verbose output
  -W<warning>             Enable the specified warning
  -Xclang <arg>           Pass <arg> to the clang compiler

The /clang: Option

When clang-cl is run with a set of /clang:<arg> options, it will gather all of the <arg> arguments and process them as if they were passed to the clang driver. This mechanism allows you to pass flags that are not exposed in the clang-cl options or flags that have a different meaning when passed to the clang driver. Regardless of where they appear in the command line, the /clang: arguments are treated as if they were passed at the end of the clang-cl command line.

The /Zc:dllexportInlines- Option

This causes the class-level dllexport and dllimport attributes to not apply to inline member functions, as they otherwise would. For example, in the code below S::foo() would normally be defined and exported by the DLL, but when using the /Zc:dllexportInlines- flag it is not:

struct __declspec(dllexport) S {
  void foo() {}
}

This has the benefit that the compiler doesn’t need to emit a definition of S::foo() in every translation unit where the declaration is included, as it would otherwise do to ensure there’s a definition in the DLL even if it’s not used there. If the declaration occurs in a header file that’s widely used, this can save significant compilation time and output size. It also reduces the number of functions exported by the DLL similarly to what -fvisibility-inlines-hidden does for shared objects on ELF and Mach-O. Since the function declaration comes with an inline definition, users of the library can use that definition directly instead of importing it from the DLL.

Note that the Microsoft Visual C++ compiler does not support this option, and if code in a DLL is compiled with /Zc:dllexportInlines-, the code using the DLL must be compiled in the same way so that it doesn’t attempt to dllimport the inline member functions. The reverse scenario should generally work though: a DLL compiled without this flag (such as a system library compiled with Visual C++) can be referenced from code compiled using the flag, meaning that the referencing code will use the inline definitions instead of importing them from the DLL.

Also note that like when using -fvisibility-inlines-hidden, the address of S::foo() will be different inside and outside the DLL, breaking the C/C++ standard requirement that functions have a unique address.

The flag does not apply to explicit class template instantiation definitions or declarations, as those are typically used to explicitly provide a single definition in a DLL, (dllexported instantiation definition) or to signal that the definition is available elsewhere (dllimport instantiation declaration). It also doesn’t apply to inline members with static local variables, to ensure that the same instance of the variable is used inside and outside the DLL.

Using this flag can cause problems when inline functions that would otherwise be dllexported refer to internal symbols of a DLL. For example:

void internal();

struct __declspec(dllimport) S {
  void foo() { internal(); }
}

Normally, references to S::foo() would use the definition in the DLL from which it was exported, and which presumably also has the definition of internal(). However, when using /Zc:dllexportInlines-, the inline definition of S::foo() is used directly, resulting in a link error since internal() is not available. Even worse, if there is an inline definition of internal() containing a static local variable, we will now refer to a different instance of that variable than in the DLL:

inline int internal() { static int x; return x++; }

struct __declspec(dllimport) S {
  int foo() { return internal(); }
}

This could lead to very subtle bugs. Using -fvisibility-inlines-hidden can lead to the same issue. To avoid it in this case, make S::foo() or internal() non-inline, or mark them dllimport/dllexport explicitly.

Finding Clang runtime libraries

clang-cl supports several features that require runtime library support:

  • Address Sanitizer (ASan): -fsanitize=address

  • Undefined Behavior Sanitizer (UBSan): -fsanitize=undefined

  • Code coverage: -fprofile-instr-generate -fcoverage-mapping

  • Profile Guided Optimization (PGO): -fprofile-generate

  • Certain math operations (int128 division) require the builtins library

In order to use these features, the user must link the right runtime libraries into their program. These libraries are distributed alongside Clang in the library resource directory. Clang searches for the resource directory by searching relative to the Clang executable. For example, if LLVM is installed in C:\Program Files\LLVM, then the profile runtime library will be located at the path C:\Program Files\LLVM\lib\clang\11.0.0\lib\windows\clang_rt.profile-x86_64.lib.

For UBSan, PGO, and coverage, Clang will emit object files that auto-link the appropriate runtime library, but the user generally needs to help the linker (whether it is lld-link.exe or MSVC link.exe) find the library resource directory. Using the example installation above, this would mean passing /LIBPATH:C:\Program Files\LLVM\lib\clang\11.0.0\lib\windows to the linker. If the user links the program with the clang or clang-cl drivers, the driver will pass this flag for them.

The auto-linking can be disabled with -fno-rtlib-defaultlib. If that flag is used, pass the complete flag to required libraries as described for ASan below.

If the linker cannot find the appropriate library, it will emit an error like this:

$ clang-cl -c -fsanitize=undefined t.cpp

$ lld-link t.obj -dll
lld-link: error: could not open 'clang_rt.ubsan_standalone-x86_64.lib': no such file or directory
lld-link: error: could not open 'clang_rt.ubsan_standalone_cxx-x86_64.lib': no such file or directory

$ link t.obj -dll -nologo
LINK : fatal error LNK1104: cannot open file 'clang_rt.ubsan_standalone-x86_64.lib'

To fix the error, add the appropriate /libpath: flag to the link line.

For ASan, as of this writing, the user is also responsible for linking against the correct ASan libraries.

If the user is using the dynamic CRT (/MD), then they should add clang_rt.asan_dynamic-x86_64.lib to the link line as a regular input. For other architectures, replace x86_64 with the appropriate name here and below.

If the user is using the static CRT (/MT), then different runtimes are used to produce DLLs and EXEs. To link a DLL, pass clang_rt.asan_dll_thunk-x86_64.lib. To link an EXE, pass -wholearchive:clang_rt.asan-x86_64.lib.

Windows System Headers and Library Lookup

clang-cl uses a set of different approaches to locate the right system libraries to link against when building code. The Windows environment uses libraries from three distinct sources:

  1. Windows SDK

  2. UCRT (Universal C Runtime)

  3. Visual C++ Tools (VCRuntime)

The Windows SDK provides the import libraries and headers required to build programs against the Windows system packages. Underlying the Windows SDK is the UCRT, the universal C runtime.

This difference is best illustrated by the various headers that one would find in the different categories. The WinSDK would contain headers such as WinSock2.h which is part of the Windows API surface, providing the Windows socketing interfaces for networking. UCRT provides the C library headers, including e.g. stdio.h. Finally, the Visual C++ tools provides the underlying Visual C++ Runtime headers such as stdint.h or crtdefs.h.

There are various controls that allow the user control over where clang-cl will locate these headers. The default behaviour for the Windows SDK and UCRT is as follows:

  1. Consult the command line.

    Anything the user specifies is always given precedence. The following extensions are part of the clang-cl toolset:

    • /winsysroot:

    The /winsysroot: is used as an equivalent to -sysroot on Unix environments. It allows the control of an alternate location to be treated as a system root. When specified, it will be used as the root where the Windows Kits is located.

    • /winsdkversion:

    • /winsdkdir:

    If /winsysroot: is not specified, the /winsdkdir: argument is consulted as a location to identify where the Windows SDK is located. Contrary to /winsysroot:, /winsdkdir: is expected to be the complete path rather than a root to locate Windows Kits.

    The /winsdkversion: flag allows the user to specify a version identifier for the SDK to prefer. When this is specified, no additional validation is performed and this version is preferred. If the version is not specified, the highest detected version number will be used.

  2. Consult the environment.

    TODO: This is not yet implemented.

    This will consult the environment variables:

    • WindowsSdkDir

    • UCRTVersion

  3. Fallback to the registry.

    If no arguments are used to indicate where the SDK is present, and the compiler is running on Windows, the registry is consulted to locate the installation.

The Visual C++ Toolset has a slightly more elaborate mechanism for detection.

  1. Consult the command line.

    • /winsysroot:

    The /winsysroot: is used as an equivalent to -sysroot on Unix environments. It allows the control of an alternate location to be treated as a system root. When specified, it will be used as the root where the VC directory is located.

    • /vctoolsdir:

    • /vctoolsversion:

    If /winsysroot: is not specified, the /vctoolsdir: argument is consulted as a location to identify where the Visual C++ Tools are located. If /vctoolsversion: is specified, that version is preferred, otherwise, the highest version detected is used.

  2. Consult the environment.

    • /external:[VARIABLE]

      This specifies a user identified environment variable which is treated as a path delimiter (;) separated list of paths to map into -imsvc arguments which are treated as -isystem.

    • INCLUDE and EXTERNAL_INCLUDE

      The path delimiter (;) separated list of paths will be mapped to -imsvc arguments which are treated as -isystem.

    • LIB (indirectly)

      The linker link.exe or lld-link.exe will honour the environment variable LIB which is a path delimiter (;) set of paths to consult for the import libraries to use when linking the final target.

    The following environment variables will be consulted and used to form paths to validate and load content from as appropriate:

    • VCToolsInstallDir

    • VCINSTALLDIR

    • Path

  3. Consult ISetupConfiguration [Windows Only]

    Assuming that the toolchain is built with USE_MSVC_SETUP_API defined and is running on Windows, the Visual Studio COM interface ISetupConfiguration will be used to locate the installation of the MSVC toolset.

  4. Fallback to the registry [DEPRECATED]

    The registry information is used to help locate the installation as a final fallback. This is only possible for pre-VS2017 installations and is considered deprecated.

Restrictions and Limitations compared to Clang

Strict Aliasing

Strict aliasing (TBAA) is always off by default in clang-cl. Whereas in clang, strict aliasing is turned on by default for all optimization levels.

To enable LLVM optimizations based on strict aliasing rules (e.g., optimizations based on type of expressions in C/C++), user will need to explicitly pass -fstrict-aliasing to clang-cl.