Also specify the -stackvar option when using any of the parallelization options, including -autopar. The -stackvar option may provide better performance when using -autopar because it may allow the optimizer to detect additional opportunities for parallelization. See the description of the -stackvar option for information on how to set the sizes for the main thread stack and for the slave thread stacks.
Auto Data 3.40 66
Download File: https://tinurll.com/2vH1ob
Mixing static Fortran runtime system libraries with dynamic Fortran runtime system libraries is not recommended and can result in linker errors or silent data corruption. Always link with the latest shared dynamic Fortran runtime system libraries.
The data layout effect is that of the -f flag: double- and quad-precision data in COMMON blocks and EQUIVALENCE classes are laid out in memory along their "natural" alignment, which is on 8-byte boundaries (or 16-byte boundaries for quad-precision when compiling for 64-bit platforms with -m64.) The default alignment in COMMON blocks and standard-conforming numeric sequence derived types is on 4-byte boundaries.
Note also that a program compiled with -fast may show good performance and accurate results with some data sets, but not with others. Avoid compiling with -fast those programs that depend on particular properties of floating-point arithmetic.
Because -fast invokes -dalign, -fns, -fsimple=2, programs compiled with -fast can result in nonstandard floating-point arithmetic, nonstandard alignment of data, and nonstandard ordering of expression evaluation. These selections might not be appropriate for most programs.
Enables automatic generation of floating-point fused multiply-add instructions. -fma=none disables generation of these instructions. -fma=fused allows the compiler to attempt to find opportunities to improve the performance of the code by using floating-point fused multiply-add instructions.
Do not automatically link with any system or language library; that is do not pass any default -lx options on to ld. The normal behavior is to link system libraries into the executables automatically, without the user specifying them on the command line.
This option inserts padding between arrays or character variables, if they are static local and not initialized, or if they are in common blocks. The extra padding positions the data to make better use of cache. In either case, the arrays or character variables can not be equivalenced.
Profiles generated by using prof(1) or gprof(1) on 64-bit Oracle Solaris platforms or just gprof on 32-bit Oracle Solaris platforms include approximate user CPU times. These times are derived from PC sample data (see pcsample(2)) for routines in the main executable and routines in shared libraries specified as linker arguments when the executable is linked. Other shared libraries (libraries opened after process startup using dlopen(3C)) are not profiled.
Use this option flag carefully. It could cause interface problems when a subroutine or function expecting a REAL*4 argument is called with a REAL*4 constant that gets promoted to REAL*8. It could also cause problems with programs reading unformatted data files written by an unformatted write with REAL*4 constants on the I/O list.
OpenMP data-sharing attributes-clauses are problematic. For example, declaring a variable "shared" whose accesses in an OpenMP parallel region may cause data race, or declaring a variable "private" whose value in a parallel region is used after the parallel region.
Stack overflows, especially in multithreaded applications with large arrays allocated on the stack, can cause silent data corruption in neighboring thread stacks. Compile all routines with -xcheck=stkovf if stack overflow is suspected. But note that compiling with this flag does not guarantee that all stack overflow situations will be detected since they could occur in routines not compiled with this flag.
Files not explicitly declared by this flag are assumed to be native files. For example, compiling with -xfilebyteorder=little4:zork.out declares zork.out to be a little-endian 32-bit x86 file with a 4-byte maximum data alignment. All other files in the program are native files.
The declared types in data records shared between big-endian and little-endian platforms must have the same sizes. For example, a file produced by a SPARC executable compiled with -xtypemap=integer:64,real:64,double:128 cannot be read by an x86 executable compiled with -xtypemap=integer:64,real:64,double:64 since the default double precision data types will have different sizes. (However, note that starting with the release of Sun Studio 12 Update 1, double:128 is accepted on x64 processors.)
An I/O operation with an entire UNION/MAP data object on a file specified as non-native will result in a runtime I/O error. You can only execute I/O operations using the individual members of the MAP (and not an entire VAX record containing the UNION/MAP) on non-native files.
With -xhwcprof enabled, the compiler generates information that helps tools associate profiled load and store instructions with the data-types and structure members (in conjunction with symbolic information produced with -g) to which they refer. It associates profile data with the data space of the target, rather than the instruction space, and provides insight into behavior that is not easily obtained from only instruction profiling.
If -xjobs is not specified, the default behavior is -xjobs=auto. This can be overridden by adding -xjobs=n to the command line. Multiple instances of -xjobs on the command line override each other until the right-most instance is reached.
The -xmemalign flag allows the user to specify the maximum memory alignment of data to be assumed by the compiler for those indeterminate situations. It also specifies the error behavior at runtime when a misaligned memory access does take place.
The prefetch latency is the hardware delay between the execution of a prefetch instruction and the time the data being prefetched is available in the cache. The compiler assumes a prefetch latency value when determining how far apart to place a prefetch instruction and the load or store instruction that uses the prefetched data.
For applications with datasets that reside entirely within the external cache, you may be able to obtain better performance by decreasing the prefetch latency values. To decrease the values, use a factor that is less than 1.
With -xprefetch, -xprefetch=auto, and -xprefetch=yes, the compiler is free to insert prefetch instructions into the code it generates. This may result in a performance improvement on architectures that support prefetch.
Options such as -xdepend can affect the aggressiveness of computing the indirect prefetch candidates and therefore the aggressiveness of the automatic indirect prefetch insertion due to better memory alias disambiguation information.
This option causes execution frequency data to be collected and saved during execution, then the data can be used in subsequent runs to improve performance. Profile collection is safe for multithreaded applications. That is, profiling a program that does its own multitasking ( -mt ) produces accurate results. This option is only valid when you specify -xO2 or greater level of optimization. If compilation and linking are performed in separate steps, the same -xprofile option must appear on the link step as well as the compile step.
-xMerge, -ztext, and -xprofile=collect should not be used together. While -xMerge forces statically initialized data into read-only storage, -ztext prohibits position-dependent symbol relocations in read-only storage, and -xprofile=collect generates statically initialized, position-dependent symbol relocations in writable storage.
The profile directory name profdir, if specified, is the pathname of the directory where profile data are to be stored when a program or shared library containing the profiled object code is executed. If the profdir pathname is not absolute, it is interpreted relative to the current working directory when the program is compiled with the option -xprofile=use:profdir.
If these environment variables are not set, the profile data is written to the directory profdir.profile in the current directory, where profdir is the name of the executable or the name specified in the -xprofile=collect:profdir flag. -xprofile does not append. profile to profdir if profdir already ends in .profile. If you run the program several times, the execution frequency data accumulates in the profdir.profile directory; that is output from prior executions is not lost.
If the optional profdir argument is specified, the compiler will create a profile directory at the specified location The data stored in the profile directory can be used either by tcov(1) or by the compiler with -xprofile=use:profdir. If the optional profdir pathname is omitted, a profile directory will be created when the profiled program is executed. The data stored in the profile directory can only be used by tcov(1). The location of the profile directory can be controlled using environment variables SUN_PROFDATA and SUN_PROFDATA_DIR.
If the location specified by profdir is not an absolute pathname, it is interpreted at compilation time relative to the current working directory at the time of compilation. If profdir is specified for any object file, the same location must be specified for all object files in the same program. The directory whose location is specified by profdir must be accessible from all machines where the profiled program is to be executed. The profile directory should not be deleted until its contents are no longer needed, because data stored there by the compiler cannot be restored except by recompilation.
Example [1]: if object files for one or more programs are compiled with -xprofile=tcov:/test/profdata, a directory named /test/profdata.profile will be created by the compiler and used to store data describing the profiled object files. The same directory will also be used at execution time to store execution data associated with the profiled object files. 2ff7e9595c
Comments