[LBNL]

Berkeley UPC - Unified Parallel C

(A joint project of LBNL and UC Berkeley)
[UCB]

Home
Downloads
Documentation
Bugs
Publications
Demos
Contact
Internal

Allocating, Initializing, and Referring to Static User Data in the Berkeley UPC Compiler

Version 1.1, October 22, 2002

Jason Duell < >


This document describes the interface between the UPC compiler and the UPC runtime for handling static user data (both shared and unshared) in UPC programs.

Within this document, 'static' user data means 'not dynamically allocated' (i.e., not allocated on the stack, nor with malloc(), upc_all_alloc(), or any other memory allocation function). All of a user's global and static variables in the regular C sense are static user data for the purposes of this document.

Allocating and initializing static data in UPC is much more challenging than in regular C, where the linker simply gathers up the static data defined in various object files and places it in an executable along with any initial values (all of which are, in C, known at link time at the latest). In UPC we cannot always know the size, location, or initial value of a variable at link time, and thus support from the runtime layer is needed to properly allocate and initialize static data. In the Berkeley UPC compiler, the mechanisms we use to set up static data also require us to also refer to it specially during program execution.

The following example shows data definitions from two UPC files (and a shared '.uph' header file) that are part of the same program--we will use this example to illustrate the steps that need to be taken with static UPC data. [Working UPC and C files for all the code shown in this web page can be found in the 'tests/foo_bar' subdirectory of the UPC Runtime distribution].

foobar.uph
/**************************************
 * Unshared global variables 
 **************************************/

extern int quux;


/**************************************
 * Shared global variables 
 **************************************/

/* Note: absence of 'extern' means these variables will be tentatively
 * declared in every file that #includes this one.  */
shared int foo;
shared int bar;

foo.upc
#include "foobar.uph"

/**************************************
 * Shared variables 
 **************************************/

shared int foo = 3; /* explicit definition overrides tentative declaration in
		       foobar.uph */

shared int *shared pbar = &bar;
      

/**************************************
 * Unshared variables 
 **************************************/

int quux;

/**************************************
 * Functions
 **************************************/

double gethandynumber() {
    /**********************************
     * Unshared static variable
     **********************************/
    static double suspects[] = { 3.14159, 2.71828 };

    assert (quux == 0 || quux == 1);

    return suspects[quux]; 
}

extern double do_sum();

int main(int argc, char **argv)
{
    gethandynumber();
    do_sum();
}

bar.upc
#include "foobar.uph"

      
int * pquux = &quux;
shared int * pfoo = &foo;


double do_sum() {
    /**********************************
     * Shared static variable
     **********************************/
    static shared [3] double messy[16][4*THREADS] 
	= { { 1, 2, 3, 4, 5 } };
    double total;
    int i, j;

    for (i = 0; i < 16; i++)
	for (j = 0; j < THREADS; j++)
	    total += messy[i][j];
    
    return total;
}

As we can see, there are two types of data in a user's UPC program that we have to deal with: shared variables, which all UPC threads can see, and unshared variables, which are visible only to a single UPC thread. Note that 'pfoo' in bar.upc is NOT a shared variable: it is a local pointer, which happens to point to the shared integer type (it is of course not a 'normal' pointer, since more information is needed to point to a shared variable than an address. But this is a separate issue from whether it is itself shared or unshared). On the other hand, 'pbar' in foo.upc is a shared variable: it is a shared pointer to a shared integer. Also note that we've made the situation tricky by placing some of our pointers (pfoo and pquux) in a different file than the variable they are initialized to point to: in a regular C program, the linker handles resolving all such addresses, but in the UPC case things are not so simple...

In the Berkeley UPC compiler, .upc files are translated into .c files that have had all UPC specific constructs translated into C code. Below are two hand-translated .c files that should be similar to those that UPC compiler emits. Don't try to understand them all at first glance (especially the initialization code at the bottom of each file): the remainder of this document will go over each element in turn.

foo.c [source file ]
#include <upcr.h>

/************************************************
 * declarations from foobar.uph 
 ************************************************/

extern int quux; /* no special syntax needed for extern definition of a
		    thread-local variable */

/* definitions of shared variables replaced with proxy pointers with the same
 * name */
upcr_pshared_ptr_t foo;		/* tentative definitions */
upcr_pshared_ptr_t bar;


/************************************************
 * Proxy pointers to shared variables
 ************************************************/

/* explicit definition overrides above tentative definition */
upcr_pshared_ptr_t foo = UPCR_INITIALIZED_PSHARED; 

upcr_pshared_ptr_t pbar = UPCR_INITIALIZED_PSHARED;


/************************************************
 * Thread-local variables 
 ************************************************/

int 
UPCR_TLD_DEFINE_TENTATIVE(quux, 4);

/* Use typedef for array. Also, move static to global scope, 
 * and mangle name to avoid name collisions.  */
typedef double _type_suspects_MANGLED[2];

_type_suspects_MANGLED 
UPCR_TLD_DEFINE(suspects_MANGLED, 8) = { 3.14159, 2.71828 };


/**************************************
 * Functions
 **************************************/

double gethandynumber() {
    UPCR_BEGIN_FUNCTION();
    /* declaration of static 'suspects' moved to global, 
     * unstatic scope */

    assert( *((int*)UPCR_TLD_ADDR(quux)) == 0
	 || *((int*)UPCR_TLD_ADDR(quux)) == 1);

    return ((double*)UPCR_TLD_ADDR(suspects_MANGLED))[*((int*)UPCR_TLD_ADDR(quux))];
}

extern double do_sum();

/************************************************
 * UPC compiler must rename 'main' to 'user_main'
 ************************************************/

int user_main(int argc, char **argv)
{
    UPCR_BEGIN_FUNCTION();
    double trouble;
    int checkval;
    upcr_pshared_ptr_t ptmp;

    gethandynumber();
    do_sum();

    return 0;
}

/************************************************
 * Allocate/init shared variables.
 * -- run once on each node
 ************************************************/

void UPCRI_ALLOC_foo_MANGLE123 (void)
{
    UPCR_BEGIN_FUNCTION();

    upcr_startup_pshalloc_t pinfos[] = {
	{ &foo,  sizeof(int), 1, 0  },
	{ &pbar, sizeof(upcr_pshared_ptr_t), 1, 0 },
	{ &bar,  sizeof(int), 1, 0 }
    };

    /* Allocate shared data */
    upcr_startup_pshalloc(pinfos, 
	sizeof(pinfos)/sizeof(upcr_startup_pshalloc_t));

}

/************************************************
 * Initialization function for TLD variables
 * -- run once per pthread per node.
 ************************************************/

void UPCRI_INIT_foo_MANGLE123 (void)
{
    UPCR_BEGIN_FUNCTION();

    /*************************************************
     * Initialize shared data 
     *************************************************/

    /* Explicit initializations of variables living only 
     * on UPC thread 0 
     */
    if (upcr_mythread() == 0) {
       *((int*)upcr_pshared_to_local(foo)) = 3;
       *((upcr_pshared_ptr_t*)upcr_pshared_to_local(pbar)) = bar;
    } 

    /* No striped arrays to initialize in this file */

    /*************************************************
     * Initialize thread-local data
     *************************************************/

    /* Both quux and suspects_MANGLED_123 are initialized 
     * satisfactorily by runtime: no special logic needed here
     */
}


bar.c [source file ]
#include <upcr.h>

/************************************************
 * declarations from foobar.uph 
 ************************************************/

extern int quux; /* no special syntax needed for extern definition of a
		    thread-local variable */

/* definitions of shared variables replaced with proxy pointers with the same
 * name */
upcr_pshared_ptr_t foo;		/* tentative definitions */
upcr_pshared_ptr_t bar;


/************************************************
 * Thread-local variables 
 ************************************************/

/* pquux requires special treatment in initialization function 
 * below, since '&quux' is different for different threads 
 */
int *
UPCR_TLD_DEFINE(pquux, 4) = &quux;

/* pfoo also requires special treatment, since the memory it points to isn't
 * allocated until startup
 */
upcr_pshared_ptr_t
UPCR_TLD_DEFINE(pfoo, 4) = UPCR_INITIALIZED_PSHARED;


/***********************************************************************
 * Function-scope static shared variables 
 * - Proxy pointer used, and is static, but promoted to file scope
 * - Function name added to proxy pointer name to avoid name conflicts 
 ***********************************************************************/
static upcr_shared_ptr_t do_sum_messy = UPCR_INITIALIZED_SHARED;



/**************************************
 * Functions
 **************************************/

double do_sum() {
    UPCR_BEGIN_FUNCTION();
    /* 'messy' moved to file scope, so init function can see it */
    double total;
    int i, j;

    for (i = 0; i < 16; i++)
	for (j = 0; j < upcr_threads(); j++) {
	    double tmp;
	    upcr_get_shared(&tmp, do_sum_messy, 
			    sizeof(double)*(16*i + j), 
			    sizeof(double));
	    total += tmp;				     
	}
    return total;
}


/************************************************
* Startup allocation function
************************************************/

void UPCRI_ALLOC_bar_MANGLE123 (void)
{
    UPCR_BEGIN_FUNCTION();

    upcr_startup_pshalloc_t pinfos[] = {
	{ &foo, sizeof(int), 1, 0 },
	{ &bar, sizeof(int), 1, 0 }
    };

    upcr_startup_shalloc_t infos[] = {
	{ &do_sum_messy, 3*sizeof(double), 16*4*sizeof(double), 1 }
    };

    /* Allocate shared data */
    upcr_startup_pshalloc(pinfos,
	sizeof(pinfos) / sizeof(upcr_startup_pshalloc_t));
    upcr_startup_shalloc(infos,
	sizeof(infos) / sizeof(upcr_startup_shalloc_t));
}

/************************************************
* Startup initialization function 
************************************************/

void UPCRI_INIT_bar_MANGLE123 (void)
{
    UPCR_BEGIN_FUNCTION();

    /*************************************************
     * Initialize shared data 
     *************************************************/

    /* No thread0-specific shared initializations in this file */

    /* Have each UPC thread initialize its part of striped array */
    {	
	double init_messy[1][5] = { { 1, 2, 3, 4, 5 } };
	upcr_startup_arrayinit_diminfo_t init_messy_info[] = {
	    { 1, 16, 0 },
	    { 5, 4,  1 }
	};
	upcr_startup_initarray(do_sum_messy, init_messy, 
			       init_messy_info, 2, 
			       sizeof(double), 3);
    }

    /*************************************************
     * Initialize thread-local data
     *************************************************/

    (*((int**)UPCR_TLD_ADDR(pquux))) = UPCR_TLD_ADDR(quux);
    (*((upcr_pshared_ptr_t*)UPCR_TLD_ADDR(pfoo))) = foo;
}

A brief propaedeutic digression: "tentative definitions"

The ANSI/ISO C specification refers to variable definitions like 'int bar;' (i.e. those that do not use 'extern', yet do not provide an initial value) as 'tentative definitions' (see section 6.9.2). Multiple tentative definitions-- including multiple appearances in the same file--are allowed in a program without error (in contrast, multiple initialized definitions are always illegal, even if the same value is used for the initialization: placing multiple 'int foo = 0;' statements in a program will cause a linker error). Multiple tentative definitions must be converted into a single variable in the final application. If the variable was initialized somewhere in the application, that value must be used. Otherwise, it must an initial value of 0. For example, the declarations of both 'foo' and 'bar' in foobar.uph are tentative--but foo winds up being initialized to '3' in foo.upc, while 'bar' is never initialized, and must be set to 0.

Support for properly converting multiple tentative definitions into a single variable requires special support from the linker (compilers cannot know when they see 'int foo;' whether the variable will be initialized in a different file).

Duplicate tentative definitions are rare in real code, and typically show up only in older C code. The 'extern' keyword is now typically used to avoid multiple definitions. However, since the UPC specification states that UPC officially follows the ANSI/ISO C specification except where explicitly noted otherwise, a UPC compiler ought to handle them. This specification contains a fair amount of logic dedicated specifically to handling tentative definitions correctly (although one of our two alternatives for handling unshared global UPC variables declared by the user does not currently support them completely, as explained later).

Definition and Use of Static Shared Variables

Since GASNet does not guarantee that any particular range of virtual addresses in a UPC program will be addressable by the network, the UPC compiler and runtime cannot allocate any actual shared memory until program startup. To implement user defined static shared variables, an indirection strategy must be used, in which each static shared variable is represented by a "proxy" unshared pointer to the shared memory, which is allocated at startup. The following list explains the steps that need to be taken to make this scheme work.
  1. Definitions of shared static variables replaced with unshared 'proxy' pointers

    Wherever the compiler sees a definition of a static shared variable, it instead defines the variable as type upcr_shared_ptr_t or upcr_pshared_ptr. For the rest of this document, these are referred to as 'proxy pointers'.

    Note: the 'phaseless' upcr_pshared_ptr_t type is used (to save space and/or make address calculation easier) when the variable is either a scalar value that will live only on thread 0, or an array that either exists entirely on a single thread (i.e. is indefinitely blocked), or which uses the default UPC blocking of one element per block.

        /*** UPC code ***/
    
        shared int foo = 3;
        shared int bar;
        double do_sum() {
            static shared [3] double messy[16][4*THREADS] = { ... };
        }
    
        /*** Translated C code ***/
    
        upcr_pshared_ptr_t foo = UPCR_INITIALIZED_PSHARED;
        upcr_pshared_ptr_t bar;
        static upcr_shared_ptr_t do_sum_messy = UPCR_INITIALIZED_SHARED;
        
    A couple points are worth noting here.

  2. References to shared variables replaced with use of proxy pointers and UPC runtime functions

    All references to shared static variables are performed via their proxy variables and the appropriate UPC runtime functions: so the array access of 'messy' in 'do_sum' is transformed into
        double tmp;
        upcr_get_shared(&tmp, do_sum_messy, 
                        sizeof(double)*(16*i + j), 
                        sizeof(double));
        total += tmp;				     
    
    Note that any optimizations performed by the compiler to avoid, schedule, or coalesce network traffic are performed above the level of the UPC runtime--the code here, for instance, might be altered by an enterprising compiler to use a single block copy per thread.
  3. Allocation/initializations performed in per-file init/allocation functions

    As described later in this document, each .c file generated by the UPC compiler must contain two initialization functions: one to dynamically allocate the shared memory for all shared static data items defined in the file, and one which performs initializations of both shared and unshared user variables defined in the file (samples of these functions are provided in the foo.c and bar.c files, named with the prefixes UPCRI_ALLOC_ and UPCRI_INIT_, respectively).

    The allocation function must contain upcr_startup_{p}shalloc_t structs with the allocation information for each proxy pointer defined in the file:

        upcr_startup_pshalloc_t pinfos[] = {
    	{ &foo, sizeof(int), 1, 0 },
    	{ &bar, sizeof(int), 1, 0 }
        };
    
        upcr_startup_shalloc_t infos[] = {
    	{ &do_sum_messy, 3*sizeof(double), 16*4*sizeof(double), 1 }
        };
    
        /* Allocate shared data */
        upcr_startup_pshalloc(pinfos,
    	sizeof(pinfos) / sizeof(upcr_startup_pshalloc_t));
        upcr_startup_shalloc(infos,
    	sizeof(infos) / sizeof(upcr_startup_shalloc_t));
    
    A call to upcr_startup_{p}shalloc() is then made to actually allocate the shared memory for each proxy pointer (and spread the information about it to all of the node/threads in the UPC job). The function takes the address of the proxy pointer, the size and number of blocks of shared memory to allocate, and a flag indicating if the number of blocks should be multiplied by THREADS. The function also performs a bzero() on the data if it was never initialized by the user (which can be determined by noting whether the proxy pointer's initial value was UPCR_INITIALIZED_{P}SHARED or not).

    In the initialization function for the file, all shared data that was initialized by the user must be assigned the correct values. Scalar shared values will all have affinity to thread 0, and so only that thread should run the code that sets the values. Here, for instance, is the relevant code from foo.c:

        /* Explicit initializations of variables living only 
         * on UPC thread 0 
         */
        if (upcr_mythread() == 0) {
          *((int*)upcr_pshared_to_local(foo)) = 3;
          *((upcr_shared_ptr_t*)upcr_pshared_to_local(pbar)) = bar;
        } 
    
    [Note that casting to local pointers is not the only way to achieve this--it was done here since calling upcr_put_pshared() would have first required storing the '3' in a temporary variable, and the author was feeling lazy. Compilers may generate any code that correctly does the job].

    For arrays that are striped across UPC threads, initialization is trickier, and a helper function called upcr_startup_assignarray() function is provided. It takes a pointer to a local array from which the initial values for the shared array will be taken, and a set of information for each dimension of the arrays. Each thread initializes only the portion of the array which has affinity to it, to avoid unneeded network traffic. If the local array is not as large as the shared array, the remainder of shared array is filled with 0s.

        double init_messy[1][5] = { { 1, 2, 3, 4, 5 } };
        upcr_startup_arrayinit_diminfo_t init_messy_info[] = {
            { 1, 16, 0 },
            { 5, 4,  1 }
        };
        upcr_startup_initarray(do_sum_messy, init_messy, 
                               init_messy_info, 2, 
                               sizeof(double), 3);
    
    See the UPC Runtime Specification for more details on the parameters and behaviors of these functions.

Definition and Use of Static Unshared Variables (Thread-Local Data)

If single-threaded processes are used to run a UPC job, the user's global/static unshared variables can simply be declared and referenced as regular C global/static variables. But if pthreads are used to implement multiple UPC threads within a process, this no longer would work, since the UPC specification mandates that each UPC thread has its own copy of such data. Instead we need to make per-thread copies of all such data items (i.e. they must be made 'thread-local').

There are various ways to transform global/static data into thread-local data. The Berkeley UPC compiler supports two methods: a 'global struct' approach, and a 'tld section' approach. Both strategies cause all such data across all files to be coalesced into a single region, a copy of which is made for each thread. References to thread-local variables are then transformed into offsets into the current thread's region.

Each strategy has its disadvantages: the 'global struct' approach occasionally requires all .c files in a UPC application to be recompiled, and uses more memory at runtime. The 'tld section' strategy requires compiler and linker behaviors that are not portable across different C compilers.

While this discussion is concerned specifically with the case when the UPC compiler is generating C output, the strategies (especially the 'tld section' approach) should also be relevant to UPC compilers that generate straight to object code.

Interface to thread-local data declarations/definitions/references

The interface between the UPC compiler and the UPC runtime for declaring, defining, and referring to unshared static variables is the same for both threaded and unthreaded targets, and for both TLD implementation strategies. This allows the UPC compiler to generate the same C code regardless of the compilation environment.
  1. Declarations made with 'extern' are noted as TLD, but are otherwise unchanged

    Declarations that are made with the 'extern' keyword do not need to be transformed by the UPC compiler, and should be output unchanged in the output C file:
        extern int defined_somewhere_else;    /* same in both .upc and .c output */
    
    There is an important exception to this rule--unshared pointers to shared data still need to be transformed into upcr_shared_ptr_t's:
        extern shared int *pint;            /* in .upc */
    
        extern upcr_pshared_ptr_t pint;      /* in .c output */
    
    Although the UPC compiler need not transform a 'extern' declaration itself, it does need to note the fact that the data in question is thread-local since such items are not referred to in the normal way, as we will see below.

  2. Definitions with initial value use UPC_TLD_DEFINE macro

    Definitions that provide an initial value for a variable use the UPCR_TLD_DEFINE() macro:
        int mcfoobar = 999;                 /* in .upc */
    
        /* in .c output */
        int UPCR_TLD_DEFINE(mcfoobar, 4) = 999;
    
    The macro takes the name and size (in bytes) of the variable. The full type of the definition must come before the macro, so
        int natural[3] = { 1, 2, 3};
        void (*int_taker)(int) = &print_int;
    
    cannot be transformed into
        int UPCR_TLD_DEFINE(natural)[3] = {1, 2, 3}
        void (*(UPCR_TLD_DEFINE(int_taker, 4))(int) = &print_int;
    
    Instead the UPC compiler must declare typedefs for array and function pointer definitions:
        typedef int _type_natural[3];
        _type_natural UPCR_TLD_DEFINE(natural, 12) = { 1, 2, 3 };
    
        typedef void (*_type_int_taker)(int);
        _type_int_taker UPCR_TLD_DEFINE(int_taker, 4) = &print_int;
    
    Finally, static unshared definitions must be promoted to regular (unstatic) type and global scope, and when this is done, their names must be mangled to avoid any name collisions with other global variables that may exist in other files (the 'suspects' array in foo.upc is an example of such a variable). Such mangling should be done in a deterministic fashion, so that the name of the variable is not changed across compilations unnecessarily (it is OK for the name to change whenever the set of names/sizes of other global unshared data change, but it should not change otherwise).

  3. Tentative definitions use UPCR_TLD_DEFINE_TENTATIVE macro

    Tentative declarations of unshared global/static variables must use the UPCR_TLD_DEFINE_TENTATIVE macro. Thus
        int quux;
    
    at file scope in foo.upc becomes
        int UPCR_TLD_DEFINE_TENTATIVE(quux, 4);
    
    in foo.c.

    The macro otherwise works identically to UPCR_TLD_DEFINE.

  4. Regular "C" variables declared/defined in .h/.c files are not treated as thread-local

    UPC must support programs that wish to link to standard C system libraries, or to other C code that the user may provide. The user should not call the UPC compiler on .c files, but .upc/.uph files will need to be able to #include .h (and perhaps .c) files and still link correctly.

    To link and operate correctly with regular C libraries, UPC must not treat data it sees in .c/.h files as thread-local variables: instead it must treat them as regular global variables. Variables are recognized as being external C variables if they are declared/defined in a #included .h or .c file.

    Of course, for this strategy to work with a pthreaded UPC process, all linked C code must be thread-safe. UPC applications which need to use non-thread-safe C code or libraries should compile and run their UPC code as single-threaded executables.

  5. References to TLD performed via UPCR_TLD_ADDR macro

    Once a variable has been flagged as thread-local by the compiler, all references to it must be done via the UPCR_TLD_ADDR() macro, which returns the address of the thread's copy of the data. The address is returned as a void *, so it must be cast to the correct type, and dereferenced as needed. For instance,
        assert (quux == 0 || quux == 1);
    
    in foo.upc must be converted into
        assert( *((int*)UPCR_TLD_ADDR(quux)) == 0
    	 || *((int*)UPCR_TLD_ADDR(quux)) == 1);
     
    in foo.c.

  6. Complex initializations performed in file-specific init functions

    Unlike with shared data, most TLD will be correctly initialized at startup by the UPC runtime without any specific action needed on the part of the UPC compiler. However, initializations that involve addresses of other TLD, such as
            int *pquux = &quux;
    
    must be handled specially (since the address of quux will be different on different pthreads). Local pointers to shared data also require special treatment:
            shared int pfoo =  &foo;
    
    cannot be correctly assigned until the shared memory for 'foo' is allocated at startup. The UPC compiler must recognize all such special cases, and perform the appropriate assignments in each file's initialization function (information on the per-file allocation/initialization functions is provided later in this document). The above two definitions in bar.upc, for instance, cause the following special logic in bar.c's initialization function:
        (*((int**)UPCR_TLD_ADDR(pquux))) = UPCR_TLD_ADDR(quux);
        (*((upcr_shared_ptr_t*)UPCR_TLD_ADDR(pfoo))) = foo;
    

 

The 'tld section' approach to implementing thread-local data

The key to the 'tld section' approach is that the linker is used to coalesce all TLD into a special separate data segment of the final executable. A copy of this segment can then be made in memory at startup for each additional pthread on a UPC node (process), and references to TLD items can be transformed into offsets into a thread's copy of the TLD section.

The compiler directives used in the explanation below are all specific to the GNU GCC compiler. They also may not work (even with GCC) if the target machine does not support the ELF object format. Other C compilers may use different compiler/linker directives to achieve the same effect, or may not support the strategy at all. For this reason UPC compilers which target C code as their output may find it easier (and more portable) to use the 'global struct' strategy. Authors of UPC compilers which directly produce object code, however, will probably find the 'tld section' approach more natural within a compiler context.

Finally, as specified here, the tld section approach does not support multiple tentative definitions of the same UPC variable in multiple files (it does support it for variables defined in external C header files). It has not yet been determined if full support for tentative definitions is achievable under the tld section approach--at a minimum it appears that a custom linker script would need to be written to make them work. In the worst case it could certainly be done by modifying the linker itself.

  1. Static unshared definitions are coalesced into thread-local linkage section

    The UPC compiler will coalesce unshared static data items into a separate section (let's call it '.upc_tld') of each object file. It does this by expanding the UPCR_TLD_DEFINE() macro to a compiler command that causes the variable to be placed in a special linkage section: When GCC/ELF is the target compiler/format, for instance, the UPC runtime header file simply arrange for the appropriate __attribute__ modifier to be placed in the declaration (the size parameter of the macro is simply discarded): so
        int UPCR_TLD_DEFINE(jrandomvariable, 4) = 9;
    
    becomes
        int jrandomvariable __attribute__((section(".upc_tld"))) = 9;
    
    Any ELF-compatible linker will automatically coalesce the '.upc_tld' sections from the various object files into a single, contiguous '.upc_tld' section in the executable.

    The party pooper: tentative definitions

    You will note that we do not mention the UPCR_TLD_DEFINE_TENTATIVE macro here. This is because we have not yet figured out a way to get it to work correctly.

    In regular C tentative definitions are placed in a special 'common' section of .o files. Multiple definitions of the same variable are permitted to exist in the various object files that are linked to form an executable, so long as at most one such variable is in an initialized data section. At link time the linker examines each variable defined in the common sections of the objects to be linked: if an initialized value exists, it is used, otherwise the object is created in the 'BSS' (i.e. it is created with an initial value of 0).

    The gcc documentation states that the __attribute__((section)) directive only works with initialized values, and is ignored for uninitialized variables. In actuality, at least in recent gcc versions, the directive does not get ignored, and instead causes the variable to be put in the desired section with an initial value of 0. This, alas, is not sufficient, since if the same variable appears in multiple object files (even with the same initial 0 value), the linker declares a duplicate symbol error. One can avoid linker errors by causing the UPCR_TLD_DEFINE_TENTATIVE macro to use __attribute__(weak)), but this in turn causes the 'section' attribute to be ignored, so the variable will not be made thread-local.

    It may be possible to have the UPCR_TLD_DEFINE_TENTATIVE use a different section name (ex: .upc_tld_common), and then somehow write a linker script that will treat that section with the common section's semantics at link time, but is not known if this will work (the author's several pleas for help on the gnu.gcc Usenet group have gone unanswered).

    Another alternative may simply be to ban the use of multiple tentative definitions within UPC code, while supporting them for extern "C" code. Multiple tentative definitions can always be trivially avoided without any change in program semantics via the addition of an 'extern', and programmers writing new UPC code are unlikely to even notice the absence of full support for tentative definitions (C++, for instance, does not use tentative definitions--'int foo;' is equivalent to 'int foo = 0;'--but few programmers are even aware of this difference). Old C libraries may place tentative definitions in their header files, but since such code is treated as 'extern C' by the UPC compiler (and hence will not be converted into thread-local data), such definitions will still be handled correctly. If it is decided that support for multiple tentative unshared UPC variables is not needed, UPCR_TLD_DEFINE_TENTATIVE can simply be #defined to UPCR_TLD_DEFINE, and single tentative definitions will work correctly.

  2. Length and address of the static unshared segment stored in known locations

    The UPC compiler will arrange to have the starting address and length of the .upc_tld segment written into two 'well-known' variables that are visible to the UPC runtime. This will probably need to be done in a linker script.

  3. Copy of linkage section made for each new UPC thread at startup

    At startup, for each UPC thread besides the first, the UPC runtime will allocate a copy of the .upc_tld section. A simple memcopy will be done to move the initial values in the .upc_tld section into each of these copies, and then the per-file initialization functions (described elsewhere in this document) will be run to perform any special initializations that need to be done. Each thread will be given a pointer to its thread-local static unshared data area, and will use it for all static unshared data references.
  4. Conversion of references into dynamic lookups

    At compile time, the compiler will translate references of thread-local data into a dynamic calculation that takes the offset of the referred-to variable within the .upc_tld section, and adds it to the address of the thread's copy of this area. C code targets will do this with an inline function call:
        UPCR_TLD_ADDR(foo)
    
    which will return the equivalent of
        (tld_addrs[MYTHREAD] + ( ((uintptr_t)&foo) - upc_tld_start))
    
    cast to a void pointer.

    Since the 'tld_addrs[THREAD] - upc_tld_start' portion can be done only once, at startup, and then stored as a separate 'tld_offset[MYTHREAD]' variable, the cost of a lookup can be optimized to

            tld_offset[MYTHREAD] + (unintptr_t)&foo)
    
    On most architectures, this should translate into a single indexed load instruction (assuming the value of tld_offset[MYTHREAD] is cached in a register).

 

The 'global struct' approach to implementing thread-local data

The global struct approach uses a regular C structure to coalesce thread-local variables. All unshared global/static variables in all files are converted into members of this structure, whose definition is made visible to each .c file. At startup, a copy of the struct is made for each thread and initialized, and references to thread-local variables are converted into member accesses in the struct.
  1. UPCR_TLD_DEFINE definitions grepped out into .tld files

    When foo.upc is compiled into foo.c by the UPC compiler, the script that invokes the compiler will also cause a second program to be run after the .c file is generated. This program will be a grep-like tool that finds all the UPCR_TLD_DEFINE (and UPCR_TLD_DEFINE_TENTATIVE) macro invocations in the .c file, sorts them (in a deterministic order which is otherwise unspecified), and dumps them out into a foo.tld file. This file will contain only lines that look like
        UPCR_TLD_DEFINE(suspects_MANGLED, 8)
        UPCR_TLD_DEFINE(quux, 4)
    
    Note that any UPCR_TLD_DEFINE_TENTATIVE definitions are transformed into regular UPCR_TLD_DEFINEs in the .tld file (we do not need to distinguish between them here). Also, any duplicate definitions are discarded. Also note the lack of semicolons in the .tld file. Finally, the grep-like script will only overwrite an existing .tld data if its contents are different (this will only happen if a variable has been added/deleted/renamed, or its size has changed).
  2. UPCR_TLD_DEFINE macros in .c file expands into regular variable definition

    When the global struct approach is used, the regular UPC runtime header files will define the UPCR_TLD_DEFINE{_TENTATIVE} macros to simply expand to the 'name' parameter. Thus the backend C compiler will see a regular definition, like:
        int quux;
    
    These regular global variables serve several purposes. First, they store the initial value (if any) for the definition. Secondly, they will cause the linker to catch any errors from the user initializing the value in multiple files. Third, the linker will handle tentative definitions of these variables correctly. These variables are otherwise unused in the final executable, and this is what makes the global struct approach consume more memory than the tld section approach (which can use the initial coalesced linker section of thread-local data as thread 0's section, only making copies for further threads).
  3. upcr_global_tld.h creates struct from UPCR_TLD_DEFINE macros

    When the global struct approach is in use, upcr.h (which is #included by all .c files generated by the UPC compiler) will at some point #include upcr_global_tld.h. This file contains the declaration of the global structure that holds all thread-local data:
        #define UPCR_TLD_DEFINE(name, size)  char name[size];
        struct upcr_tld {
        #include "upcr_global_tld.tld"
        };
        #undef UPCR_TLD_DEFINE
    
        /* at some point later in upcr.h or a file it includes... */
        #define UPCR_TLD_DEFINE(name, size) name
    
    All variables are declared as the same type--arrays of char. This is done because it is virtually impossible to assemble the full set of type information that would needed to use the real types of the variables as they are declared in various scopes and .c files (the same type name may legally be used in different files/scopes to refer to different typedefs/structs. Correct ordering of type declarations is also difficult). Since the UPCR_TLD_ADDR macro returns a void * (and the compiler will always know what type to cast it to), this is not a problem. Alignment issues can be solved by sorting the definitions in the global tld file by size, and/or by padding the sizes of variables passed to UPCR_TLD_DEFINE{_TENTATIVE}.
  4. upcr_global_tld.c creates initialization function from UPCR_TLD_DEFINE macros

    The upcr_global_tld.c file uses the UPCR_TLD_DEFINE macros to provide an initialization function for the global struct:
        #include <string.h>
        #include "upcr_global_tld.h"
    
        #undef UPCR_TLD_DEFINE
        #define UPCR_TLD_DEFINE(name, size) extern int name;
        #include "upcr_global_tld.tld"
    
        void upcri_startup_init_tld(struct upcr_tld *tld)
        {
        #undef UPCR_TLD_DEFINE
        #define UPCR_TLD_DEFINE(name, size) memcpy(&tld->name, &name, size);
        #include "upcr_global_tld.tld"
        }
    
    The function uses the values of the global variables left in the .c files as the source for initial values. Any initializations for which this simple memcpy is not sufficient must be handled in the special per-file initialization functions.
  5. UPC linker script sets up and performs "hidden make" to build executable

    The UPC compiler and linker will present the familiar "compile .c files into .o objects, then link the .o files with a linker" interface to users.

    When the 'tld section' approach is used (or the UPC executable will run as a single-threaded process), the UPC compiler will also invoke the backend C compiler on its output .c files, the resulting .o files seen by the user will be regular C object files, and the UPC linker wrapper will simply send them directly to the regular C linker. When the 'global struct' approach is in use, however, UPC .o files will actually be copies of the .c files output by the UPC compiler: all compilation by the back-end C compiler will be done at link time, since this is the only time that enough information is available to know the full layout of the global thread-local data structure.

    The fact that all C compilation occurs at link time in the global struct approach does not mean, however, that every intermediate .c file in a UPC application needs to be recompiled every time the application is linked. The fact that users typically link an application with the same set of files repeatedly can be exploited by the UPC linker to avoid needless recompilations. Under this scheme, UPC .o files will only be recompiled when the .o file itself has been changed (presumably because the user has modified and recompiled its parent .upc file), or when the size or layout of the global tld struct has changed (in which case all the UPC .o files in the application will need to be recompiled).

    This optimization is performed via the following steps:

    1. When the UPC linker is invoked, it examines the list of files it was given, and looks in a "hidden" build directory (perhaps 'upc-build') for a Makefile that corresponds to this list of files (note: the linker is able to tell between UPC .o files--which contain C source code--and regular C .o files that have been created by another compiler. Only UPC .o files count for purposes of determining if the file list is 'different').

    2. If a Makefile setup for this exact list already exists, the script jumps to the final 'make' step below.

    3. If the Makefile exists but is for a different list of files, 'make clean' is executed, and the Makefile recreated as if it did not exist (see next step).

    4. If the Makefile does not exist, the script will first ensure that the hidden build directory exists (creating it if necessary), and that two files, 'upcr_global_tld.h' and 'upcr_global_tld.c' are in it (they are copied from the UPC compiler's installation directory if needed). A Makefile with the following characteristics is then built and placed in the build directory:

      • The Makefile will construct a list of '.upo' files corresponding to the the UPC .o files passed to the linker. The target application will depend on OBJS, which will consist of these .upo files, plus upcr_global_tld.o (any regular C .o files or libraries the user passed along to the linker script will become EXTRA_OBJS, which are not involved in any dependencies, but which will be linked into the executable). The rule for making .upo files will first transform a .o file into a .c file (so the C compiler will recognize it), then compile it regularly into the .upo file.
      • Each .upo file in OBJS will depend on both its .o file, and on "upcr_global_tld.h".
      • A list of .tld files will be created from the UPC .o files passed to the linker, and upcr_global_tld.h will depend on these. The rule for building upcr_global_tld.h will run a script that coalesces all the various .tld files into a single upcr_global_tld.tld file (and updates upcr_global_tld.h with a simple 'touch'). Sorting of entries in the global_tld file is not required (note: one might want to base the ordering on program profile data for optimal cache performance), but only one copy of any duplicate entries (which can result from the same tentative definition appearing in different .upc files) may be kept.

    5. "make" will be invoked, and the program will be linked and built. The Makefile's dependency structure will allow future relinking of the same set of UPC files to often proceed without requiring recompilation of all the .c files--if a single UPC file is changed and recompiled, the other .c files in the directory will only be recompiled if the user changed the names or sizes of any thread-specific data in the UPC file (i.e. if they changed the names or sizes of any unshared global/static variables). This should allow for a much faster compile/run/modify debug cycle in most cases.

    Note: The 'hidden' build directory and all the files it in will not be deleted after the link is complete. Thus, users will need to explicitly delete the 'upc-build' directory in their 'make clean' commands--it will never be deleted automatically for them.

 

UPC Startup Framework and Per-file Allocation/Initialization Functions

As this document has shown, the UPC compiler needs to generate a fair amount of allocation and initialization code that is run at startup. The content of this startup code depends on the user's UPC code, and so it cannot all be boilerplate library code. Also, since separate compilation is used, the logic cannot even be centralized into a single function in the final executable--instead each file must contain the startup logic that is needed to support the code in that file.

The question then becomes how to arrange to call all of these functions at startup, how to name them in such a way that they do not collide in the symbol namespace, and how to determine the order in which they are called. The Berkeley UPC compiler takes the following approach to these issues:


Home
Downloads
Documentation
Bugs
Publications
Demos
Contact
Internal

This page last modified on Thursday, 31-Aug-2017 17:19:14 PDT