These options control various sorts of optimizations.
Without any optimization option, the compiler's goal is to reduce the cost of compilation and to make debugging produce the expected results. Statements are independent: if you stop the program with a breakpoint between statements, you can then assign a new value to any variable or change the program counter to any other statement in the function and get exactly the results you expect from the source code.
Turning on optimization flags makes the compiler attempt to improve the performance and/or code size at the expense of compilation time and possibly the ability to debug the program.
The compiler performs optimization based on the knowledge it has of the program. Compiling multiple files at once to a single output file mode allows the compiler to use information gained from all of the files when compiling each of them.
Not all optimizations are controlled directly by a flag. Only optimizations that have a flag are listed in this section.
Most optimizations are only enabled if an -O level is set on the command line. Otherwise they are disabled, even if individual optimization flags are specified.
Depending on the target and how GCC was configured, a slightly different set of optimizations may be enabled at each -O level than those listed here. You can invoke GCC with -Q --help=optimizers to find out the exact set of optimizations that are enabled at each level. See Overall Options, for examples.
-O
-O1
With -O, the compiler tries to reduce code size and execution time, without performing any optimizations that take a great deal of compilation time.
-O turns on the following optimization flags:
-fauto-inc-dec -fcompare-elim -fcprop-registers -fdce -fdefer-pop -fdelayed-branch -fdse -fguess-branch-probability -fif-conversion2 -fif-conversion -fipa-pure-const -fipa-profile -fipa-reference -fmerge-constants -fsplit-wide-types -ftree-bit-ccp -ftree-builtin-call-dce -ftree-ccp -ftree-ch -ftree-copyrename -ftree-dce -ftree-dominator-opts -ftree-dse -ftree-forwprop -ftree-fre -ftree-phiprop -ftree-slsr -ftree-sra -ftree-pta -ftree-ter -funit-at-a-time
-O also turns on -fomit-frame-pointer on machines
where doing so does not interfere with debugging.
-O2
-O2 turns on all optimization flags specified by -O. It also turns on the following optimization flags:
-fthread-jumps -falign-functions -falign-jumps -falign-loops -falign-labels -fcaller-saves -fcrossjumping -fcse-follow-jumps -fcse-skip-blocks -fdelete-null-pointer-checks -fdevirtualize -fdevirtualize-speculatively -fexpensive-optimizations -fgcse -fgcse-lm -fhoist-adjacent-loads -finline-small-functions -findirect-inlining -fipa-sra -fisolate-erroneous-paths-dereference -foptimize-sibling-calls -fpartial-inlining -fpeephole2 -freorder-blocks -freorder-functions -frerun-cse-after-loop -fsched-interblock -fsched-spec -fschedule-insns -fschedule-insns2 -fstrict-aliasing -fstrict-overflow -ftree-switch-conversion -ftree-tail-merge -ftree-pre -ftree-vrp
Please note the warning under -fgcse about
invoking -O2 on programs that use computed gotos.
-O3
-O0
-Os
-Os disables the following optimization flags:
-falign-functions -falign-jumps -falign-loops -falign-labels -freorder-blocks -freorder-blocks-and-partition -fprefetch-loop-arrays
-Ofast
-Og
If you use multiple -O options, with or without level numbers, the last such option is the one that is effective.
Options of the form -fflag specify machine-independent flags. Most flags have both positive and negative forms; the negative form of -ffoo is -fno-foo. In the table below, only one of the forms is listed—the one you typically use. You can figure out the other form by either removing ‘no-’ or adding it.
The following options control specific optimizations. They are either activated by -O options or are related to ones that are. You can use the following flags in the rare cases when “fine-tuning” of optimizations to be performed is desired.
-fno-defer-pop
Disabled at levels -O, -O2, -O3, -Os.
-fforward-propagate
This option is enabled by default at optimization levels -O,
-O2, -O3, -Os.
-ffp-contract=
styleThe default is -ffp-contract=fast.
-fomit-frame-pointer
On some machines, such as the VAX, this flag has no effect, because
the standard calling sequence automatically handles the frame pointer
and nothing is saved by pretending it doesn't exist. The
machine-description macro FRAME_POINTER_REQUIRED
controls
whether a target machine supports this flag. See Register Usage.
Starting with GCC version 4.6, the default setting (when not optimizing for size) for 32-bit GNU/Linux x86 and 32-bit Darwin x86 targets has been changed to -fomit-frame-pointer. The default can be reverted to -fno-omit-frame-pointer by configuring GCC with the --enable-frame-pointer configure option.
Enabled at levels -O, -O2, -O3, -Os.
-foptimize-sibling-calls
Enabled at levels -O2, -O3, -Os.
-fno-inline
always_inline
attribute. This is the default when not
optimizing.
Single functions can be exempted from inlining by marking them
with the noinline
attribute.
-finline-small-functions
Enabled at level -O2.
-findirect-inlining
Enabled at level -O2.
-finline-functions
If all calls to a given function are integrated, and the function is
declared static
, then the function is normally not output as
assembler code in its own right.
Enabled at level -O3.
-finline-functions-called-once
static
functions called once for inlining into their
caller even if they are not marked inline
. If a call to a given
function is integrated, then the function is not output as assembler code
in its own right.
Enabled at levels -O1, -O2, -O3 and -Os.
-fearly-inlining
always_inline
and functions whose body seems
smaller than the function call overhead early before doing
-fprofile-generate instrumentation and real inlining pass. Doing so
makes profiling significantly cheaper and usually inlining faster on programs
having large chains of nested wrapper functions.
Enabled by default.
-fipa-sra
Enabled at levels -O2, -O3 and -Os.
-finline-limit=
nInlining is actually controlled by a number of parameters, which may be specified individually by using --param name=value. The -finline-limit=n option sets some of these parameters as follows:
max-inline-insns-single
max-inline-insns-auto
See below for a documentation of the individual parameters controlling inlining and for the defaults of these parameters.
Note: there may be no value to -finline-limit that results in default behavior.
Note: pseudo instruction represents, in this particular context, an
abstract measurement of function's size. In no way does it represent a count
of assembly instructions and as such its exact meaning might change from one
release to an another.
-fno-keep-inline-dllexport
dllexport
attribute or declspec (See Declaring Attributes of Functions.)
-fkeep-inline-functions
static
functions that are declared inline
into the object file, even if the function has been inlined into all
of its callers. This switch does not affect functions using the
extern inline
extension in GNU C90. In C++, emit any and all
inline functions into the object file.
-fkeep-static-consts
static const
when optimization isn't turned
on, even if the variables aren't referenced.
GCC enables this option by default. If you want to force the compiler to
check if a variable is referenced, regardless of whether or not
optimization is turned on, use the -fno-keep-static-consts option.
-fmerge-constants
This option is the default for optimized compilation if the assembler and linker support it. Use -fno-merge-constants to inhibit this behavior.
Enabled at levels -O, -O2, -O3, -Os.
-fmerge-all-constants
This option implies -fmerge-constants. In addition to
-fmerge-constants this considers e.g. even constant initialized
arrays or initialized constant variables with integral or floating-point
types. Languages like C or C++ require each variable, including multiple
instances of the same variable in recursive calls, to have distinct locations,
so using this option results in non-conforming
behavior.
-fmodulo-sched
-fmodulo-sched-allow-regmoves
-fno-branch-count-reg
The default is -fbranch-count-reg.
-fno-function-cse
This option results in less efficient code, but some strange hacks that alter the assembler output may be confused by the optimizations performed when this option is not used.
The default is -ffunction-cse
-fno-zero-initialized-in-bss
This option turns off this behavior because some programs explicitly rely on variables going to the data section—e.g., so that the resulting executable can find the beginning of that section and/or make assumptions based on that.
The default is -fzero-initialized-in-bss.
-fthread-jumps
Enabled at levels -O2, -O3, -Os.
-fsplit-wide-types
long
long
on a 32-bit system, split the registers apart and allocate them
independently. This normally generates better code for those types,
but may make debugging more difficult.
Enabled at levels -O, -O2, -O3,
-Os.
-fcse-follow-jumps
if
statement with an
else
clause, CSE follows the jump when the condition
tested is false.
Enabled at levels -O2, -O3, -Os.
-fcse-skip-blocks
if
statement with no else clause,
-fcse-skip-blocks causes CSE to follow the jump around the
body of the if
.
Enabled at levels -O2, -O3, -Os.
-frerun-cse-after-loop
Enabled at levels -O2, -O3, -Os.
-fgcse
Note: When compiling a program using computed gotos, a GCC extension, you may get better run-time performance if you disable the global common subexpression elimination pass by adding -fno-gcse to the command line.
Enabled at levels -O2, -O3, -Os.
-fgcse-lm
Enabled by default when -fgcse is enabled.
-fgcse-sm
Not enabled at any optimization level.
-fgcse-las
Not enabled at any optimization level.
-fgcse-after-reload
-faggressive-loop-optimizations
-funsafe-loop-optimizations
-fcrossjumping
Enabled at levels -O2, -O3, -Os.
-fauto-inc-dec
-fdce
-fdse
-fif-conversion
if-conversion2
.
Enabled at levels -O, -O2, -O3, -Os.
-fif-conversion2
Enabled at levels -O, -O2, -O3, -Os.
-fdeclone-ctor-dtor
Enabled by -Os.
-fdelete-null-pointer-checks
Note however that in some environments this assumption is not true. Use -fno-delete-null-pointer-checks to disable this optimization for programs that depend on that behavior.
Some targets, especially embedded ones, disable this option at all levels.
Otherwise it is enabled at all levels: -O0, -O1,
-O2, -O3, -Os. Passes that use the information
are enabled independently at different optimization levels.
-fdevirtualize
-findirect-inlining
) and interprocedural constant
propagation (-fipa-cp).
Enabled at levels -O2, -O3, -Os.
-fdevirtualize-speculatively
-fexpensive-optimizations
Enabled at levels -O2, -O3, -Os.
-free
Enabled for Alpha, AArch64 and x86 at levels -O2,
-O3, -Os.
-flive-range-shrinkage
-fira-algorithm=
algorithm-fira-region=
region-fira-hoist-pressure
This option is enabled at level -Os for all targets.
-fira-loop-pressure
This option is enabled at level -O3 for some targets.
-fno-ira-share-save-slots
-fno-ira-share-spill-slots
-fira-verbose=
n-fdelayed-branch
Enabled at levels -O, -O2, -O3, -Os.
-fschedule-insns
Enabled at levels -O2, -O3.
-fschedule-insns2
Enabled at levels -O2, -O3, -Os.
-fno-sched-interblock
-fno-sched-spec
-fsched-pressure
-fsched-spec-load
-fsched-spec-load-dangerous
-fsched-stalled-insns
-fsched-stalled-insns=
n-fsched-stalled-insns-dep
-fsched-stalled-insns-dep=
n-fsched2-use-superblocks
This only makes sense when scheduling after register allocation, i.e. with
-fschedule-insns2 or at -O2 or higher.
-fsched-group-heuristic
-fsched-critical-path-heuristic
-fsched-spec-insn-heuristic
-fsched-rank-heuristic
-fsched-last-insn-heuristic
-fsched-dep-count-heuristic
-freschedule-modulo-scheduled-loops
-fselective-scheduling
-fselective-scheduling2
-fsel-sched-pipelining
-fsel-sched-pipelining-outer-loops
-fshrink-wrap
-fcaller-saves
This option is always enabled by default on certain machines, usually those which have no call-preserved registers to use instead.
Enabled at levels -O2, -O3, -Os.
-fcombine-stack-adjustments
Enabled by default at -O1 and higher.
-fconserve-stack
-ftree-reassoc
-ftree-pre
-ftree-partial-pre
-ftree-forwprop
-ftree-fre
-ftree-phiprop
-fhoist-adjacent-loads
-ftree-copy-prop
-fipa-pure-const
-fipa-reference
-fipa-pta
-fipa-profile
cold
, noreturn
, static constructors or destructors) are identified. Cold
functions and loop less parts of functions executed once are then optimized for
size.
Enabled by default at -O and higher.
-fipa-cp
-fipa-cp-clone
-fisolate-erroneous-paths-dereference
-fisolate-erroneous-paths-attribute
returns_nonnull
or nonnull
attribute. Isolate those paths from the main control flow and turn the
statement with erroneous or undefined behaviour into a trap. This is not
currently enabled, but may be enabled by -O2
in the future.
-ftree-sink
-ftree-bit-ccp
-ftree-ccp
-ftree-switch-conversion
-ftree-tail-merge
-ftree-dce
-ftree-builtin-call-dce
errno
but are otherwise side-effect free. This flag is
enabled by default at -O2 and higher if -Os is not also
specified.
-ftree-dominator-opts
-ftree-dse
-ftree-ch
-ftree-loop-optimize
-ftree-loop-linear
-floop-interchange
DO J = 1, M DO I = 1, N A(J, I) = A(J, I) * C ENDDO ENDDO
loop interchange transforms the loop as if it were written:
DO I = 1, N DO J = 1, M A(J, I) = A(J, I) * C ENDDO ENDDO
which can be beneficial when N
is larger than the caches,
because in Fortran, the elements of an array are stored in memory
contiguously by column, and the original loop iterates over rows,
potentially creating at each access a cache miss. This optimization
applies to all the languages supported by GCC and is not limited to
Fortran. To use this code transformation, GCC has to be configured
with --with-ppl and --with-cloog to enable the
Graphite loop transformation infrastructure.
-floop-strip-mine
DO I = 1, N A(I) = A(I) + C ENDDO
loop strip mining transforms the loop as if it were written:
DO II = 1, N, 51 DO I = II, min (II + 50, N) A(I) = A(I) + C ENDDO ENDDO
This optimization applies to all the languages supported by GCC and is
not limited to Fortran. To use this code transformation, GCC has to
be configured with --with-ppl and --with-cloog to
enable the Graphite loop transformation infrastructure.
-floop-block
DO I = 1, N DO J = 1, M A(J, I) = B(I) + C(J) ENDDO ENDDO
loop blocking transforms the loop as if it were written:
DO II = 1, N, 51 DO JJ = 1, M, 51 DO I = II, min (II + 50, N) DO J = JJ, min (JJ + 50, M) A(J, I) = B(I) + C(J) ENDDO ENDDO ENDDO ENDDO
which can be beneficial when M
is larger than the caches,
because the innermost loop iterates over a smaller amount of data
which can be kept in the caches. This optimization applies to all the
languages supported by GCC and is not limited to Fortran. To use this
code transformation, GCC has to be configured with --with-ppl
and --with-cloog to enable the Graphite loop transformation
infrastructure.
-fgraphite-identity
-floop-nest-optimize
-floop-parallelize-all
-fcheck-data-deps
-ftree-loop-if-convert
-ftree-loop-if-convert-stores
for (i = 0; i < N; i++) if (cond) A[i] = expr;
is transformed to
for (i = 0; i < N; i++) A[i] = cond ? expr : A[i];
potentially producing data races.
-ftree-loop-distribution
DO I = 1, N A(I) = B(I) + C D(I) = E(I) * F ENDDO
is transformed to
DO I = 1, N A(I) = B(I) + C ENDDO DO I = 1, N D(I) = E(I) * F ENDDO
-ftree-loop-distribute-patterns
This pass distributes the initialization loops and generates a call to memset zero. For example, the loop
DO I = 1, N A(I) = 0 B(I) = A(I) + I ENDDO
is transformed to
DO I = 1, N A(I) = 0 ENDDO DO I = 1, N B(I) = A(I) + I ENDDO
and the initialization loop is transformed into a call to memset zero.
-ftree-loop-im
-ftree-loop-ivcanon
-fivopts
-ftree-parallelize-loops=n
-ftree-pta
-ftree-sra
-ftree-copyrename
-ftree-coalesce-inlined-vars
-ftree-coalesce-vars
-ftree-ter
-ftree-slsr
-ftree-vectorize
-ftree-loop-vectorize
-ftree-slp-vectorize
-fvect-cost-model=
modelunlimited
, dynamic
or cheap
.
With the unlimited
model the vectorized code-path is assumed
to be profitable while with the dynamic
model a runtime check
will guard the vectorized code-path to enable it only for iteration
counts that will likely execute faster than when executing the original
scalar loop. The cheap
model will disable vectorization of
loops where doing so would be cost prohibitive for example due to
required runtime checks for data dependence or alignment but otherwise
is equal to the dynamic
model.
The default cost model depends on other optimization flags and is
either dynamic
or cheap
.
-fsimd-cost-model=
modelunlimited
, dynamic
, cheap
. All values of model
have the same meaning as described in -fvect-cost-model and by
default a cost model defined with -fvect-cost-model is used.
-ftree-vrp
-ftracer
-funroll-loops
-funroll-all-loops
-fsplit-ivs-in-unroller
A combination of -fweb and CSE is often sufficient to obtain the same effect. However, that is not reliable in cases where the loop body is more complicated than a single basic block. It also does not work at all on some architectures due to restrictions in the CSE pass.
This optimization is enabled by default.
-fvariable-expansion-in-unroller
-fpartial-inlining
Enabled at level -O2.
-fpredictive-commoning
This option is enabled at level -O3.
-fprefetch-loop-arrays
This option may generate better or worse code; results are highly dependent on the structure of loops within the source code.
Disabled at level -Os.
-fno-peephole
-fno-peephole2
-fpeephole is enabled by default.
-fpeephole2 enabled at levels -O2, -O3, -Os.
-fno-guess-branch-probability
GCC uses heuristics to guess branch probabilities if they are not provided by profiling feedback (-fprofile-arcs). These heuristics are based on the control flow graph. If some branch probabilities are specified by ‘__builtin_expect’, then the heuristics are used to guess branch probabilities for the rest of the control flow graph, taking the ‘__builtin_expect’ info into account. The interactions between the heuristics and ‘__builtin_expect’ can be complex, and in some cases, it may be useful to disable the heuristics so that the effects of ‘__builtin_expect’ are easier to understand.
The default is -fguess-branch-probability at levels
-O, -O2, -O3, -Os.
-freorder-blocks
Enabled at levels -O2, -O3.
-freorder-blocks-and-partition
This optimization is automatically turned off in the presence of exception handling, for linkonce sections, for functions with a user-defined section attribute and on any architecture that does not support named sections.
Enabled for x86 at levels -O2, -O3.
-freorder-functions
.text.hot
for most frequently executed functions and
.text.unlikely
for unlikely executed functions. Reordering is done by
the linker so object file format must support named sections and linker must
place them in a reasonable way.
Also profile feedback must be available to make this option effective. See -fprofile-arcs for details.
Enabled at levels -O2, -O3, -Os.
-fstrict-aliasing
unsigned int
can alias an int
, but not a
void*
or a double
. A character type may alias any other
type.
Pay special attention to code like this:
union a_union { int i; double d; }; int f() { union a_union t; t.d = 3.0; return t.i; }
The practice of reading from a different union member than the one most recently written to (called “type-punning”) is common. Even with -fstrict-aliasing, type-punning is allowed, provided the memory is accessed through the union type. So, the code above works as expected. See Structures unions enumerations and bit-fields implementation. However, this code might not:
int f() { union a_union t; int* ip; t.d = 3.0; ip = &t.i; return *ip; }
Similarly, access by taking the address, casting the resulting pointer and dereferencing the result has undefined behavior, even if the cast uses a union type, e.g.:
int f() { double d = 3.0; return ((union a_union *) &d)->i; }
The -fstrict-aliasing option is enabled at levels
-O2, -O3, -Os.
-fstrict-overflow
i + 10 > i
is always true for
signed i
. This assumption is only valid if signed overflow is
undefined, as the expression is false if i + 10
overflows when
using twos complement arithmetic. When this option is in effect any
attempt to determine whether an operation on signed numbers
overflows must be written carefully to not actually involve overflow.
This option also allows the compiler to assume strict pointer
semantics: given a pointer to an object, if adding an offset to that
pointer does not produce a pointer to the same object, the addition is
undefined. This permits the compiler to conclude that p + u >
p
is always true for a pointer p
and unsigned integer
u
. This assumption is only valid because pointer wraparound is
undefined, as the expression is false if p + u
overflows using
twos complement arithmetic.
See also the -fwrapv option. Using -fwrapv means that integer signed overflow is fully defined: it wraps. When -fwrapv is used, there is no difference between -fstrict-overflow and -fno-strict-overflow for integers. With -fwrapv certain types of overflow are permitted. For example, if the compiler gets an overflow when doing arithmetic on constants, the overflowed value can still be used with -fwrapv, but not otherwise.
The -fstrict-overflow option is enabled at levels
-O2, -O3, -Os.
-falign-functions
-falign-functions=
n-fno-align-functions and -falign-functions=1 are equivalent and mean that functions are not aligned.
Some assemblers only support this flag when n is a power of two; in that case, it is rounded up.
If n is not specified or is zero, use a machine-dependent default.
Enabled at levels -O2, -O3.
-falign-labels
-falign-labels=
n-fno-align-labels and -falign-labels=1 are equivalent and mean that labels are not aligned.
If -falign-loops or -falign-jumps are applicable and are greater than this value, then their values are used instead.
If n is not specified or is zero, use a machine-dependent default which is very likely to be ‘1’, meaning no alignment.
Enabled at levels -O2, -O3.
-falign-loops
-falign-loops=
n-fno-align-loops and -falign-loops=1 are equivalent and mean that loops are not aligned.
If n is not specified or is zero, use a machine-dependent default.
Enabled at levels -O2, -O3.
-falign-jumps
-falign-jumps=
n-fno-align-jumps and -falign-jumps=1 are equivalent and mean that loops are not aligned.
If n is not specified or is zero, use a machine-dependent default.
Enabled at levels -O2, -O3.
-funit-at-a-time
Enabled by default.
-fno-toplevel-reorder
asm
statements. Output them in the same order that they appear in the
input file. When this option is used, unreferenced static variables
are not removed. This option is intended to support existing code
that relies on a particular ordering. For new code, it is better to
use attributes when possible.
Enabled at level -O0. When disabled explicitly, it also implies
-fno-section-anchors, which is otherwise enabled at -O0 on some
targets.
-fweb
Enabled by default with -funroll-loops.
-fwhole-program
main
and those merged by attribute externally_visible
become static functions
and in effect are optimized more aggressively by interprocedural optimizers.
This option should not be used in combination with -flto
.
Instead relying on a linker plugin should provide safer and more precise
information.
-flto[=
n]
To use the link-time optimizer, -flto and optimization options should be specified at compile time and during the final link. For example:
gcc -c -O2 -flto foo.c gcc -c -O2 -flto bar.c gcc -o myprog -flto -O2 foo.o bar.o
The first two invocations to GCC save a bytecode representation of GIMPLE into special ELF sections inside foo.o and bar.o. The final invocation reads the GIMPLE bytecode from foo.o and bar.o, merges the two files into a single internal image, and compiles the result as usual. Since both foo.o and bar.o are merged into a single image, this causes all the interprocedural analyses and optimizations in GCC to work across the two files as if they were a single one. This means, for example, that the inliner is able to inline functions in bar.o into functions in foo.o and vice-versa.
Another (simpler) way to enable link-time optimization is:
gcc -o myprog -flto -O2 foo.c bar.c
The above generates bytecode for foo.c and bar.c, merges them together into a single GIMPLE representation and optimizes them as usual to produce myprog.
The only important thing to keep in mind is that to enable link-time optimizations you need to use the GCC driver to perform the link-step. GCC then automatically performs link-time optimization if any of the objects involved were compiled with the -flto. You generally should specify the optimization options to be used for link-time optimization though GCC will try to be clever at guessing an optimization level to use from the options used at compile-time if you fail to specify one at link-time. You can always override the automatic decision to do link-time optimization at link-time by passing -fno-lto to the link command.
To make whole program optimization effective, it is necessary to make certain whole program assumptions. The compiler needs to know what functions and variables can be accessed by libraries and runtime outside of the link-time optimized unit. When supported by the linker, the linker plugin (see -fuse-linker-plugin) passes information to the compiler about used and externally visible symbols. When the linker plugin is not available, -fwhole-program should be used to allow the compiler to make these assumptions, which leads to more aggressive optimization decisions.
When -fuse-linker-plugin is not enabled then, when a file is compiled with -flto, the generated object file is larger than a regular object file because it contains GIMPLE bytecodes and the usual final code (see -ffat-lto-objects. This means that object files with LTO information can be linked as normal object files; if -fno-lto is passed to the linker, no interprocedural optimizations are applied. Note that when -fno-fat-lto-objects is enabled the compile-stage is faster but you cannot perform a regular, non-LTO link on them.
Additionally, the optimization flags used to compile individual files are not necessarily related to those used at link time. For instance,
gcc -c -O0 -ffat-lto-objects -flto foo.c gcc -c -O0 -ffat-lto-objects -flto bar.c gcc -o myprog -O3 foo.o bar.o
This produces individual object files with unoptimized assembler code, but the resulting binary myprog is optimized at -O3. If, instead, the final binary is generated with -fno-lto, then myprog is not optimized.
When producing the final binary, GCC only applies link-time optimizations to those files that contain bytecode. Therefore, you can mix and match object files and libraries with GIMPLE bytecodes and final object code. GCC automatically selects which files to optimize in LTO mode and which files to link without further processing.
There are some code generation flags preserved by GCC when generating bytecodes, as they need to be used during the final link stage. Generally options specified at link-time override those specified at compile-time.
If you do not specify an optimization level option -O at link-time then GCC will compute one based on the optimization levels used when compiling the object files. The highest optimization level will win here.
Currently, the following options and their setting are take from the first object file that explicitely specified it: -fPIC, -fpic, -fpie, -fcommon, -fexceptions, -fnon-call-exceptions, -fgnu-tm and all the -m target flags.
Certain ABI changing flags are required to match in all compilation-units and trying to override this at link-time with a conflicting value is ignored. This includes options such as -freg-struct-return and -fpcc-struct-return.
Other options such as -ffp-contract, -fno-strict-overflow, -fwrapv, -fno-trapv or -fno-strict-aliasing are passed through to the link stage and merged conservatively for conflicting translation units. Specifically -fno-strict-overflow, -fwrapv and -fno-trapv take precedence and for example -ffp-contract=off takes precedence over -ffp-contract=fast. You can override them at linke-time.
It is recommended that you compile all the files participating in the same link with the same options and also specify those options at link time.
If LTO encounters objects with C linkage declared with incompatible types in separate translation units to be linked together (undefined behavior according to ISO C99 6.2.7), a non-fatal diagnostic may be issued. The behavior is still undefined at run time. Similar diagnostics may be raised for other languages.
Another feature of LTO is that it is possible to apply interprocedural optimizations on files written in different languages:
gcc -c -flto foo.c g++ -c -flto bar.cc gfortran -c -flto baz.f90 g++ -o myprog -flto -O3 foo.o bar.o baz.o -lgfortran
Notice that the final link is done with g++ to get the C++ runtime libraries and -lgfortran is added to get the Fortran runtime libraries. In general, when mixing languages in LTO mode, you should use the same link command options as when mixing languages in a regular (non-LTO) compilation.
If object files containing GIMPLE bytecode are stored in a library archive, say
libfoo.a, it is possible to extract and use them in an LTO link if you
are using a linker with plugin support. To create static libraries suitable
for LTO, use gcc-ar and gcc-ranlib instead of ar
and ranlib
; to show the symbols of object files with GIMPLE bytecode, use
gcc-nm. Those commands require that ar, ranlib
and nm have been compiled with plugin support. At link time, use the the
flag -fuse-linker-plugin to ensure that the library participates in
the LTO optimization process:
gcc -o myprog -O2 -flto -fuse-linker-plugin a.o b.o -lfoo
With the linker plugin enabled, the linker extracts the needed GIMPLE files from libfoo.a and passes them on to the running GCC to make them part of the aggregated GIMPLE image to be optimized.
If you are not using a linker with plugin support and/or do not
enable the linker plugin, then the objects inside libfoo.a
are extracted and linked as usual, but they do not participate
in the LTO optimization process. In order to make a static library suitable
for both LTO optimization and usual linkage, compile its object files with
-flto -ffat-lto-objects
.
Link-time optimizations do not require the presence of the whole program to operate. If the program does not require any symbols to be exported, it is possible to combine -flto and -fwhole-program to allow the interprocedural optimizers to use more aggressive assumptions which may lead to improved optimization opportunities. Use of -fwhole-program is not needed when linker plugin is active (see -fuse-linker-plugin).
The current implementation of LTO makes no attempt to generate bytecode that is portable between different types of hosts. The bytecode files are versioned and there is a strict version check, so bytecode files generated in one version of GCC will not work with an older or newer version of GCC.
Link-time optimization does not work well with generation of debugging information. Combining -flto with -g is currently experimental and expected to produce unexpected results.
If you specify the optional n, the optimization and code generation done at link time is executed in parallel using n parallel jobs by utilizing an installed make program. The environment variable MAKE may be used to override the program used. The default value for n is 1.
You can also specify -flto=jobserver to use GNU make's
job server mode to determine the number of parallel jobs. This
is useful when the Makefile calling GCC is already executing in parallel.
You must prepend a ‘+’ to the command recipe in the parent Makefile
for this to work. This option likely only works if MAKE is
GNU make.
-flto-partition=
alg1to1
to specify a partitioning mirroring
the original source files or balanced
to specify partitioning
into equally sized chunks (whenever possible) or max
to create
new partition for every symbol where possible. Specifying none
as an algorithm disables partitioning and streaming completely.
The default value is balanced
. While 1to1
can be used
as an workaround for various code ordering issues, the max
partitioning is intended for internal testing only.
-flto-compression-level=
n-flto-report
Disabled by default.
-flto-report-wpa
-fuse-linker-plugin
This option enables the extraction of object files with GIMPLE bytecode out
of library archives. This improves the quality of optimization by exposing
more code to the link-time optimizer. This information specifies what
symbols can be accessed externally (by non-LTO object or during dynamic
linking). Resulting code quality improvements on binaries (and shared
libraries that use hidden visibility) are similar to -fwhole-program
.
See -flto for a description of the effect of this flag and how to
use it.
This option is enabled by default when LTO support in GCC is enabled
and GCC was configured for use with
a linker supporting plugins (GNU ld 2.21 or newer or gold).
-ffat-lto-objects
-fno-fat-lto-objects improves compilation time over plain LTO, but requires the complete toolchain to be aware of LTO. It requires a linker with linker plugin support for basic functionality. Additionally, nm, ar and ranlib need to support linker plugins to allow a full-featured build environment (capable of building static libraries etc). GCC provides the gcc-ar, gcc-nm, gcc-ranlib wrappers to pass the right options to these tools. With non fat LTO makefiles need to be modified to use them.
The default is -fno-fat-lto-objects on targets with linker plugin
support.
-fcompare-elim
This pass only applies to certain targets that cannot explicitly represent the comparison operation before register allocation is complete.
Enabled at levels -O, -O2, -O3, -Os.
-fuse-ld=bfd
-fuse-ld=gold
-fcprop-registers
Enabled at levels -O, -O2, -O3, -Os.
-fprofile-correction
-fprofile-dir=
path-fprofile-generate
-fprofile-generate=
pathThe following options are enabled: -fprofile-arcs
, -fprofile-values
, -fvpt
.
If path is specified, GCC looks at the path to find
the profile feedback data files. See -fprofile-dir.
-fprofile-use
-fprofile-use=
pathThe following options are enabled: -fbranch-probabilities
, -fvpt
,
-funroll-loops
, -fpeel-loops
, -ftracer
, -ftree-vectorize
,
ftree-loop-distribute-patterns
By default, GCC emits an error message if the feedback profiles do not match the source code. This error can be turned into a warning by using -Wcoverage-mismatch. Note this may result in poorly optimized code.
If path is specified, GCC looks at the path to find the profile feedback data files. See -fprofile-dir.
The following options control compiler behavior regarding floating-point arithmetic. These options trade off between speed and correctness. All must be specifically enabled.
-ffloat-store
This option prevents undesirable excess precision on machines such as
the 68000 where the floating registers (of the 68881) keep more
precision than a double
is supposed to have. Similarly for the
x86 architecture. For most programs, the excess precision does only
good, but a few programs rely on the precise definition of IEEE floating
point. Use -ffloat-store for such programs, after modifying
them to store all pertinent intermediate computations into variables.
-fexcess-precision=
stylefloat
and double
types and the processor does not
support operations rounding to those types. By default,
-fexcess-precision=fast is in effect; this means that
operations are carried out in the precision of the registers and that
it is unpredictable when rounding to the types specified in the source
code takes place. When compiling C, if
-fexcess-precision=standard is specified then excess
precision follows the rules specified in ISO C99; in particular,
both casts and assignments cause values to be rounded to their
semantic types (whereas -ffloat-store only affects
assignments). This option is enabled by default for C if a strict
conformance option such as -std=c99 is used.
-fexcess-precision=standard is not implemented for languages
other than C, and has no effect if
-funsafe-math-optimizations or -ffast-math is
specified. On the x86, it also has no effect if -mfpmath=sse
or -mfpmath=sse+387 is specified; in the former case, IEEE
semantics apply without excess precision, and in the latter, rounding
is unpredictable.
-ffast-math
This option causes the preprocessor macro __FAST_MATH__
to be defined.
This option is not turned on by any -O option besides
-Ofast since it can result in incorrect output for programs
that depend on an exact implementation of IEEE or ISO rules/specifications
for math functions. It may, however, yield faster code for programs
that do not require the guarantees of these specifications.
-fno-math-errno
errno
after calling math functions that are executed
with a single instruction, e.g., sqrt
. A program that relies on
IEEE exceptions for math error handling may want to use this flag
for speed while maintaining IEEE arithmetic compatibility.
This option is not turned on by any -O option since it can result in incorrect output for programs that depend on an exact implementation of IEEE or ISO rules/specifications for math functions. It may, however, yield faster code for programs that do not require the guarantees of these specifications.
The default is -fmath-errno.
On Darwin systems, the math library never sets errno
. There is
therefore no reason for the compiler to consider the possibility that
it might, and -fno-math-errno is the default.
-funsafe-math-optimizations
This option is not turned on by any -O option since it can result in incorrect output for programs that depend on an exact implementation of IEEE or ISO rules/specifications for math functions. It may, however, yield faster code for programs that do not require the guarantees of these specifications. Enables -fno-signed-zeros, -fno-trapping-math, -fassociative-math and -freciprocal-math.
The default is -fno-unsafe-math-optimizations.
-fassociative-math
(x + 2**52) - 2**52
. May also reorder floating-point comparisons
and thus may not be used when ordered comparisons are required.
This option requires that both -fno-signed-zeros and
-fno-trapping-math be in effect. Moreover, it doesn't make
much sense with -frounding-math. For Fortran the option
is automatically enabled when both -fno-signed-zeros and
-fno-trapping-math are in effect.
The default is -fno-associative-math.
-freciprocal-math
x / y
can be replaced with x * (1/y)
, which is useful if (1/y)
is subject to common subexpression elimination. Note that this loses
precision and increases the number of flops operating on the value.
The default is -fno-reciprocal-math.
-ffinite-math-only
This option is not turned on by any -O option since it can result in incorrect output for programs that depend on an exact implementation of IEEE or ISO rules/specifications for math functions. It may, however, yield faster code for programs that do not require the guarantees of these specifications.
The default is -fno-finite-math-only.
-fno-signed-zeros
The default is -fsigned-zeros.
-fno-trapping-math
This option should never be turned on by any -O option since it can result in incorrect output for programs that depend on an exact implementation of IEEE or ISO rules/specifications for math functions.
The default is -ftrapping-math.
-frounding-math
The default is -fno-rounding-math.
This option is experimental and does not currently guarantee to
disable all GCC optimizations that are affected by rounding mode.
Future versions of GCC may provide finer control of this setting
using C99's FENV_ACCESS
pragma. This command-line option
will be used to specify the default state for FENV_ACCESS
.
-fsignaling-nans
This option causes the preprocessor macro __SUPPORT_SNAN__
to
be defined.
The default is -fno-signaling-nans.
This option is experimental and does not currently guarantee to
disable all GCC optimizations that affect signaling NaN behavior.
-fsingle-precision-constant
-fcx-limited-range
NaN
+ I*NaN
, with an attempt to rescue the situation in that case. The
default is -fno-cx-limited-range, but is enabled by
-ffast-math.
This option controls the default setting of the ISO C99
CX_LIMITED_RANGE
pragma. Nevertheless, the option applies to
all languages.
-fcx-fortran-rules
NaN
+ I*NaN
, with an attempt to rescue the situation in that case.
The default is -fno-cx-fortran-rules.
The following options control optimizations that may improve performance, but are not enabled by any -O options. This section includes experimental options that may produce broken code.
-fbranch-probabilities
With -fbranch-probabilities, GCC puts a
‘REG_BR_PROB’ note on each ‘JUMP_INSN’ and ‘CALL_INSN’.
These can be used to improve optimization. Currently, they are only
used in one place: in reorg.c, instead of guessing which path a
branch is most likely to take, the ‘REG_BR_PROB’ values are used to
exactly determine which path is taken more often.
-fprofile-values
With -fbranch-probabilities, it reads back the data gathered from profiling values of expressions for usage in optimizations.
Enabled with -fprofile-generate and -fprofile-use.
-fprofile-reorder-functions
Enabled with -fprofile-use.
-fvpt
With -fbranch-probabilities, it reads back the data gathered
and actually performs the optimizations based on them.
Currently the optimizations include specialization of division operations
using the knowledge about the value of the denominator.
-frename-registers
Enabled by default with -funroll-loops and -fpeel-loops.
-ftracer
Enabled with -fprofile-use.
-funroll-loops
Enabled with -fprofile-use.
-funroll-all-loops
-fpeel-loops
Enabled with -fprofile-use.
-fmove-loop-invariants
-funswitch-loops
-ffunction-sections
-fdata-sections
Use these options on systems where the linker can perform optimizations to improve locality of reference in the instruction space. Most systems using the ELF object format and SPARC processors running Solaris 2 have linkers with such optimizations. AIX may have these optimizations in the future.
Only use these options when there are significant benefits from doing
so. When you specify these options, the assembler and linker
create larger object and executable files and are also slower.
You cannot use gprof
on all systems if you
specify this option, and you may have problems with debugging if
you specify both this option and -g.
-fbranch-target-load-optimize
-fbranch-target-load-optimize2
-fbtr-bb-exclusive
-fstack-protector
alloca
, and
functions with buffers larger than 8 bytes. The guards are initialized
when a function is entered and then checked when the function exits.
If a guard check fails, an error message is printed and the program exits.
-fstack-protector-all
-fstack-protector-strong
-fsection-anchors
For example, the implementation of the following function foo
:
static int a, b, c; int foo (void) { return a + b + c; }
usually calculates the addresses of all three variables, but if you compile it with -fsection-anchors, it accesses the variables from a common anchor point instead. The effect is similar to the following pseudocode (which isn't valid C):
int foo (void) { register int *xr = &x; return xr[&a - &x] + xr[&b - &x] + xr[&c - &x]; }
Not all targets support this option.
--param
name=
valueThe names of specific parameters, and the meaning of the values, are tied to the internals of the compiler, and are subject to change without notice in future releases.
In each case, the value is an integer. The allowable choices for name are:
predictable-branch-outcome
max-crossjump-edges
min-crossjump-insns
max-grow-copy-bb-insns
max-goto-duplication-insns
max-delay-slot-insn-search
max-delay-slot-live-search
max-gcse-memory
max-gcse-insertion-ratio
max-pending-list-length
max-modulo-backtrack-attempts
max-inline-insns-single
max-inline-insns-auto
inline-min-speedup
large-function-insns
large-function-growth
large-unit-insns
inline-unit-growth
ipcp-unit-growth
large-stack-frame
large-stack-frame-growth
max-inline-insns-recursive
max-inline-insns-recursive-auto
For functions declared inline, --param max-inline-insns-recursive is
taken into account. For functions not declared inline, recursive inlining
happens only when -finline-functions (included in -O3) is
enabled and --param max-inline-insns-recursive-auto is used. The
default value is 450.
max-inline-recursive-depth
max-inline-recursive-depth-auto
For functions declared inline, --param max-inline-recursive-depth is
taken into account. For functions not declared inline, recursive inlining
happens only when -finline-functions (included in -O3) is
enabled and --param max-inline-recursive-depth-auto is used. The
default value is 8.
min-inline-recursive-probability
When profile feedback is available (see -fprofile-generate) the actual
recursion depth can be guessed from probability that function recurses via a
given call expression. This parameter limits inlining only to call expressions
whose probability exceeds the given threshold (in percents).
The default value is 10.
early-inlining-insns
max-early-inliner-iterations
max-early-inliner-iterations
comdat-sharing-probability
comdat-sharing-probability
min-vect-loop-bound
gcse-cost-distance-ratio
gcse-unrestricted-cost
max-hoist-depth
max-tail-merge-comparisons
max-tail-merge-iterations
max-unrolled-insns
max-average-unrolled-insns
max-unroll-times
max-peeled-insns
max-peel-times
max-peel-branches
max-completely-peeled-insns
max-completely-peel-times
max-completely-peel-loop-nest-depth
max-unswitch-insns
max-unswitch-level
lim-expensive
iv-consider-all-candidates-bound
iv-max-considered-uses
iv-always-prune-cand-set-bound
scev-max-expr-size
scev-max-expr-complexity
omega-max-vars
omega-max-geqs
omega-max-eqs
omega-max-wild-cards
omega-hash-table-size
omega-max-keys
omega-eliminate-redundant-constraints
vect-max-version-for-alignment-checks
vect-max-version-for-alias-checks
vect-max-peeling-for-alignment
max-iterations-to-track
hot-bb-count-ws-permille
hot-bb-frequency-fraction
max-predicted-iterations
builtin-expect-probability
align-threshold
align-loop-iterations
tracer-dynamic-coverage
tracer-dynamic-coverage-feedback
The tracer-dynamic-coverage-feedback is used only when profile
feedback is available. The real profiles (as opposed to statically estimated
ones) are much less balanced allowing the threshold to be larger value.
tracer-max-code-growth
tracer-min-branch-ratio
tracer-min-branch-ratio
tracer-min-branch-ratio-feedback
Similarly to tracer-dynamic-coverage two values are present, one for
compilation for profile feedback and one for compilation without. The value
for compilation with profile feedback needs to be more conservative (higher) in
order to make tracer effective.
max-cse-path-length
max-cse-insns
ggc-min-expand
The default is 30% + 70% * (RAM/1GB) with an upper bound of 100% when
RAM >= 1GB. If getrlimit
is available, the notion of “RAM” is
the smallest of actual RAM and RLIMIT_DATA
or RLIMIT_AS
. If
GCC is not able to calculate RAM on a particular platform, the lower
bound of 30% is used. Setting this parameter and
ggc-min-heapsize to zero causes a full collection to occur at
every opportunity. This is extremely slow, but can be useful for
debugging.
ggc-min-heapsize
The default is the smaller of RAM/8, RLIMIT_RSS, or a limit that
tries to ensure that RLIMIT_DATA or RLIMIT_AS are not exceeded, but
with a lower bound of 4096 (four megabytes) and an upper bound of
131072 (128 megabytes). If GCC is not able to calculate RAM on a
particular platform, the lower bound is used. Setting this parameter
very large effectively disables garbage collection. Setting this
parameter and ggc-min-expand to zero causes a full collection
to occur at every opportunity.
max-reload-search-insns
max-cselib-memory-locations
reorder-blocks-duplicate
reorder-blocks-duplicate-feedback
The reorder-block-duplicate-feedback is used only when profile
feedback is available. It may be set to higher values than
reorder-block-duplicate since information about the hot spots is more
accurate.
max-sched-ready-insns
max-sched-region-blocks
max-pipeline-region-blocks
max-sched-region-insns
max-pipeline-region-insns
min-spec-prob
max-sched-extend-regions-iters
max-sched-insn-conflict-delay
sched-spec-prob-cutoff
sched-spec-state-edge-prob-cutoff
sched-mem-true-dep-cost
selsched-max-lookahead
selsched-max-sched-times
selsched-max-insns-to-rename
sms-min-sc
max-last-value-rtl
integer-share-limit
ssp-buffer-size
min-size-for-stack-sharing
max-jump-thread-duplication-stmts
max-fields-for-field-sensitive
prefetch-latency
simultaneous-prefetches
l1-cache-line-size
l1-cache-size
l2-cache-size
min-insn-to-prefetch-ratio
prefetch-min-insn-to-mem-ratio
use-canonical-types
switch-conversion-max-branch-ratio
max-partial-antic-length
sccvn-max-scc-size
sccvn-max-alias-queries-per-access
ira-max-loops-num
ira-max-conflict-table-size
ira-loop-reserved-regs
loop-invariant-max-bbs-in-loop
loop-max-datarefs-for-datadeps
max-vartrack-size
max-vartrack-expr-depth
min-nondebug-insn-uid
ipa-sra-ptr-growth-factor
tm-max-aggregate-size
graphite-max-nb-scop-params
graphite-max-bbs-per-function
loop-block-tile-size
ipa-cp-value-list-size
ipa-cp-eval-threshold
ipa-max-agg-items
ipa-cp-loop-hint-bonus
ipa-cp-array-index-hint-bonus
lto-partitions
lto-minpartition
cxx-max-namespaces-for-diagnostic-help
sink-frequency-threshold
max-stores-to-sink
allow-load-data-races
allow-store-data-races
allow-packed-load-data-races
allow-packed-store-data-races
case-values-threshold
tree-reassoc-width
sched-pressure-algorithm
The default choice depends on the target.
max-slsr-cand-scan
asan-globals
asan-stack
asan-instrument-reads
asan-instrument-writes
asan-memintrin
asan-use-after-return
asan-instrumentation-with-call-threshold