gpoverflow(5) gpoverflow(5)
NAME
gp_overflow - GP Overflow Errors
TOPIC
This man page is intended to describe the underlying causes of and
possible solutions for overflowing the gp-relative area in the linker.
The Symptoms and a Simple Solution
GP overflow is reported by the linker with one of several error messages:
GP-relative sections overflow by 0x??? bytes. Please recompile with
a smaller -G value.
GOT overflow: please relink with -multigot.
GOT unreachable: please relink with -multigot.
These messages all indicate that the code being linked contains more than
64KB of GP-relative data area (see the next section), and attempts to
reference it with a signed 16-bit displacement from the GP register.
In the past we recommended recompiling with -xgot. Although this
generally solved the GP overflow problem it created larger, slower
programs. We developed -multigot to create multiple GOTs during linking:
-multigot prevents the GOT overflow problem while retaining the more
efficient GP relative code. -xgot will be dropped in future releases.
The preferred solution to the problem is to use the -multigot option to
ld. At some point -multigot is likely to become the default default.
Even with -multigot one must still avoid having too much GP data: if the
-G value is greater than zero it may be necessary to use a smaller -G
value (see below for definitions).
Terminology
GP GP (sometimes spelled gp) is a register in the register set of the
processor. It used in a consistent way by the software. The
register is actually an ordinary integer register (register 28) but
by software convention it is used only in a stylized way. The GP-
area is a region of memory addressable by an offset of the GP
register. See the section "Underlying Mechanisms" below.
GOT To provide Position Independent Code (PIC) which can be relocated
without modifying the text of the program (or DSO) it is sufficient
to have a data table with the actual addresses of global data (with
appropriate code generation and linking support). This data table
is constructed by ld(1) and is called the Global Offset Table (GOT).
For a further explanation of what the GOT is, see dso(5).
PIC Position Independent Code (PIC) is, as mentioned above, a method of
code generation which results in code that can be shared by multiple
users (each program must have its own data space! Code sharing and
independent data is arranged automatically by the compilation and
Page 1
gpoverflow(5) gpoverflow(5)
run-time systems). For details on PIC code generation, see the
dso(5) (for old-32bit-abi) or the "MIPSpro N32 ABI Handbook" (for
the new-32bit-abi).
-G n
This compiler option sets the -G value, n. This compiler option
affects how much data goes into the "small data" and "small bss" and
other areas (discussed in the "Underlying Mechanisms" section below
where short data segment is defined) ). The default n for old-
32bit-abi compilation (when producing PIC code) is 0. The default
for new-32bit-abi and 64bit-abi compilation is 8. To see the value
actually in use, use the -show option on the compilation line and
check the -G option actually passed to the compiler passes.
Underlying Mechanisms
The MIPS instruction set architecture provides instructions for loading
values from memory, and storing values to memory, which calculate an
effective address by adding the content of a register to a signed 16-bit
immediate value (displacement) in the instruction.
The compilers take advantage of this capability by maintaining various
frequently-used data objects in a number of small memory sections,
keeping a pointer to this area of memory in a dedicated register (GP),
and generating references to those objects as simple load/store
instructions with signed 16-bit displacements If the total size of these
memory sections exceeds 64K bytes, the displacements in some of those
instructions will likely overflow, and the linker reports a GP overflow
error.
We will refer to these small GP-relative memory sections collectively as
the short data segment below. In general, it is comprised of the Unix
sections .sdata, .sbss, The data objects placed in the short data segment
by the compiler and linker include:
Static data page pointers:
These are no longer a problem with current code generation and
linkers.
GOT pointers:
References to static data or subroutine addresses in shared code,
except for those which use page pointers mentioned above, normally
load an address from the GOT, relative to GP, and use it for the
access. By default, these address loads assume an offset from GP of
32KB or less. Using the -multigot linker option tells the linker to
arrange for as many GOTs as needed in the a.out or DSO (so the fact
each GOT is limited in size becomes irrelevant). Using the (now
deprecated) -xgot compiler option (or -TENV:large_GOT option when
compiling n32 or 64-bit programs using the MIPSPro compiler) forces
the compiler to assume full 32-bit offsets from GP, and thereby
eliminates the restriction that these pointers fit in 64KB.
Page 2
gpoverflow(5) gpoverflow(5)
Small data objects:
The compiler attempts to place small data objects in a GP-relative
section (e.g. .sdata or .sbss or .lit4 or .lit8) and to reference
them directly using a signed 16-bit displacement. The limit on the
size of objects treated this way can be changed from the default to n
bytes by using the -G n option These options must be used
consistently for all compilations in a given program or linker errors
will occur. You might get one of these errors:
"foo" used in bar.o and baz.o has different sizes.
"foo" used as gp-relative in bar.o and defined otherwise in baz.o.
"foo" cannot be turned into gp-relative.
The data objects affected by -G n are numeric literals, addresses
including those generated by the compiler, all C static variables,
and (if the "-static"flag was used) FORTRAN local variables.
Small literal objects:
The compiler also attempts to place small literal objects (e.g.
integer constants, floating point constants, addresses) in a GP-
relative section (e.g. .lit4, .lit8, or .srdata) and to reference
them directly using a signed 16-bit displacement. The limit on the
size of literal objects treated this way can be changed from the
default to n bytes by using the -G n option. The -G option must be
used consistently for all compilations in a given program, or linker
errors are likely to occur.
Each of these uses normally has a positive effect on program performance.
(In fact, performance can often be improved further by increasing the
sizes of data and literal objects allocated to the short sections.)
However, if the total size of the four categories exceeds 64KB, the
linker will report GP overflow errors, and the program must be rebuilt to
shrink the total size of GP-relative data.
Tailoring a Solution for Best Performance
The simplest solution is to use the -multigot option. Finding the best
solution requires identifying the sizes of the short data section
components (page pointers, GOT pointers, small data, and small literals),
and removing only as much as necessary. With this release this
information may be obtained from the linker by specifying the flag "-
gpinfo" directly to ld(1), or by specifying "-Wl,-gpinfo" to cc(1) or
f77(1). In other cases, this information may be obtained by specifying
the "-m -aoutkeep" flags to ld(1) (or -Wl,-m,-aoutkeep to cc(1) or CC(1)
or f77(1)) and subsequently searching the load map for the desired
sections.
Each of these components can be controlled:
Page 3
gpoverflow(5) gpoverflow(5)
Page pointers:
They are no longer an issue: compiler and linker changes now make the
the -TENV:no_page_offset option unnecessary.
GOT pointers:
These must be GP-relative, but compiling with -multigot enables the
use of multiple gp regions and the linker will place the short data
and literal sections closer to an appropriate GP so that they can
still benefit.
Small data:
Compiling with -G n for n smaller than the existing value shrinks
these sections. Use -G 0 to eliminate all GP-relative small data
references. (Note that some small data references will be replaced
by new GOT pointer references, so the savings might not be as large
as the original size of the short data sections.) Recompiling with
this option requires recompiling the entire program with it.
Small literals:
Compiling with -G n for n smaller, shrinks these sections. Use -G 0
to eliminate all GP-relative small literal references. (Note that
some small literal references will be replaced by new GOT pointer
references, so the savings might not be as large as the original size
of the short literal sections.) Recompiling with -G n requires
recompiling the entire program with it.
There is another option which may perform better than the above. The
main program and each DSO it uses each get their own short data segments.
Breaking parts of the program into separate DSOs may shrink all of the
short data segment components enough that the defaults work for some or
all of the pieces. However, be aware that subprogram calls which cross
DSO boundaries incur some extra overhead, so this solution is likely to
help performance only if the division can be made such that such
crossings are infrequent.
The -xgot option
The following is retained to help make clear what the options used with
-xgot were so you can remove them and use -multigot instead. Please do
not use -xgot. Use -multigot. -multigot requires no recompiling, just
relinking. (Any program compiled -xgot should be completely recompiled
to use -multigot.) In the future -multigot will be on by default.
The -xgot rules suggested for the earlier compiler releases were as
follows:
For n32 or 64-bit programs, recompile everything in one's program with
the options:
-avoidgpoverflow
Page 4
gpoverflow(5) gpoverflow(5)
this eliminates the problem. This flag is a synonym for
-G 0 -xgot -TENV:nopageoffset
Programs compiled for the old 32bit abi recompile with
-G 0 -xgot
-xgot is likely to have a significant negative impact on the program's
performance.
BUGS
ld earlier than IRIX 6.2 produced one page pointer per 64KB page of
static data, whether or not that page is ever referenced. As a result,
256MB of static data would fill the GP-relative area with page pointers,
leaving no space for the other GP-relative memory sections. More than
that would normally cause linker errors unless the -TENV:no_page_offset
option was used for compilation. The 6.2 and later linkers are much more
efficient in its data GOT page production making the option
-TENV:no_page_offset unnecessary.
A workaround for this problem was to compile all large data blocks from
files which contain no programs of any kind, and link them into a
separate data-only dso. Since it contains no code, there will be no
references to the overflowed GOT. This dso can then be linked against
when building the main program. Since references to it are always via the
name of the symbol (or the name of the common in FORTRAN), the data will
be referenced correctly, but since the data is defined in a dso, space
will not be allocated in the main program.
For example, suppose we wanted to define an array of 1,000,000,000
floating point values. We could then write the following source file in
C:
float bigarray[1000000000];
And subsequently compile it and link it thus:
cc -c bigarray.c ld -shared bigarray.o -o bigarray.so
Then we would link the main program thus:
cc <objects> bigarray.so <libraries>
To run it we must either install bigarray.so into /usr/lib, or else tell
the runtime linker rld(1) where to find it. One way of doing this is
setting the environment variable LD_LIBRARY_PATH to the pathname of the
directory in which bigarray.so can be found.
FORTRAN programmers use the source file
Page 5
gpoverflow(5) gpoverflow(5)
BLOCKDATA DUMMY
COMMON/mycom/bigarray(1000000000)
END
Compile it with f77(1), rather than with cc(1), the other steps are the
same.
SEE ALSO
cc(1), f77(1), ld(1), dso(5), MIPSpro N32 ABI Handbook
Page 6