changes

体系机构仿真

		DESCRIPTION OF CHANGES MADE TO PRODUCE
	      THE MULTIPROCESSOR VERSION OF SIMPLESCALAR
		FROM THE ORIGINAL UNIPROCESSOR RELEASE

			   Naraig Manjikian
	     Dept. of Electrical and Computer Engineering
			  Queen's University
		  Kingston, Ontario, Canada K7L 3N6

			     30 May 2000


The multiprocessor version of SimpleScalar that is encompassed in this
software package is adapted from release 2.0 of uniprocessor SimpleScalar.

Several existing files were modified, and several new files were added.
Furthermore, many files that appear in the original uniprocessor release
have been excluded in this multiprocessor version.

This document outlines the exclusions, additions, and changes in the
set of files that constitute the multiprocessor version of SimpleScalar.

----------------------------------------------------------------------

EXCLUDED FILES THE FROM ORIGINAL UNIPROCESSOR RELEASE

bitmap.h
bpred.c
bpred.h
cache.c
cache.h
dlite.c
dlite.h
eventq.c
eventq.h
ptrace.c
ptrace.h
resource.c
resource.h
sim-bpred.c
sim-cache.c
sim-cheetah.c
sim-fast.c
sim-outorder.c
sim-profile.c
sim-safe.c

----------------------------------------------------------------------

NEW INCLUDE FILES

>>> graphics.h

definitions for rudimentary X windows code for visualization

>>> mpsyscall.h

definitions for multiprocessor-specific #defines for system call codes;
 this file is included by a modified version of syscall.h, which is in
 turn included by syscall.c; new code in syscall.c uses the #defines in
 mpsyscall.h
 ***NOTE*** that mpsyscall.h is _also_ included by an assembly-language
   file, hence it should only contain #defines for system call codes

>>> ssmp.h

declarations for the run-time library for multiprocessing;
 ***NOTE*** that this header is _not_ used to compile the simulators;
   it is used only when compiling multiprocessor application code that
   will be simulated (a separate makefile must be used to compile
   applications)

>>> c.m4.ss

macro definitions for PARMACS and used with the m4 macro preprocessor;
 ***NOTE*** that this file is used only when building application code
   to be simulated; it is _not_ used to compile the simulators

----------------------------------------------------------------------

MODIFICATIONS OF EXISTING INCLUDE FILES

>>> regs.h

added extra dimension of MAX_PROCS to declarations of the following
 variables: regs_R, regs_F, regs_HI, regs_LO, regs_FCC, regs_PC

modified declaration of regs_init() to accept three parameters:
 pid, initial stack pointer value, initial PC value (entry point)

modified declaration of regs_dump() to accept additional pid parameter

>>> ss.h

added #define for MAX_PROCS, the maximum number of processors
 (currently 32)

>>> stats.h

added new item sc_uintarray in enumeration stat_class_t

added new element for_uintarray in union within struct stat_stat_t

added declaration of new function stat_reg_uintarray()
 to register an unsigned integer array statistical variable

>>> syscall.h

added #include for mpsyscall.h immediately after SS_SYS_??? #defines

modified declaration of ss_syscall() to accept additional pid parameter

>>> ss.def

modified definitions and uses of the following functions
 to include a pid parameter:
  - InstMULT()
  - InstMULTU()
  - InstSRA()
  - InstSRAV()

----------------------------------------------------------------------

NEW SOURCE FILES

>>> graphics.c

contains rudimentary code to create an X window and perform
 simple drawing operations

>>> mpcache.c

contains code to model two-level direct-mapped caches in a
 multiprocessor, with extensive statistics gathered on hits, misses,
 and coherence activity; includes conditional code for graphical
 visualization of coherence (enabled with a #define symbol)

>>> sim-mpfast.c

(this file is adapted from sim-fast.c in the original uniprocessor
 release; the list below outlines the differences between
 sim-mpfast.c and sim-fast.c)

added definitions for variables to maintain per-processor
 instruction counts and count for number of cycles

added declaration for array of flags indicating active processors

added declaration for count of created processes

modified body of sim_reg_stats() to include multiprocessor statistics

modified body of sim_aux_stats() to report multiprocessor statistics

modified #defines that access registers to include pid

modified #define for SYSCALL() to include pid

throughout body of sim_main(), removed *all* code related
 to USE_JUMP_TABLE

in sim_main(), added definition of local pid variable

in sim_main(), initialized count of created processes to 1
 and marked the first processor as active

in main 'while' loop within sim_main(), adder inner 'for' loop that
 scans through created processes, skipping inactive ones,
 and fetches/interprets a single instruction; all register accesses
 use local pid variable

>>> sim-mpcache.c

(this file is extends sim-mpfast.c above, hence it also has differences
 with sim-fast.c from the original uniprocessor release;
 the list below outlines *additional* changes relative to sim-mpfast.c)

added #include "mpcache.h"

added definitions for variables that reflect cache configuration that
 is set with command line options

for graphics, added definitions for variables for command line options

added macro for log base 2 calculation

modified body of sim_reg_options() to register variables related to
 command line options

modified body of sim_check_options() to verify cache configuration
 from command line options

modified body of sim_init() to call initialization code in mpcache.c

in the inner 'for' loop in sim_main(), added call to function in
 mpcache.c that traces memory accesses and models cache hit/miss
 as well as coherence; the call to the function is made _before_
 a load/store instruction is interpreted because for instructions
 of the form "lw $reg,0($reg)" the base register is modified, and
 the original memory address will be lost

>>> libssmp.S

defines assembly-language routines in the SimpleScalar portable
 instruction set architecture (PISA); these routines constitute the
 run-time library that is linked with multiprocessor application code
 that will be simulated; this file includes mpsyscall.h described above;
 ***NOTE*** that this file is _not_ used to generate the simulators;
  a _separate_ makefile is used to generate PISA code for these routines

----------------------------------------------------------------------

MODIFICATIONS OF EXISTING SOURCE FILES

>>> main.c

removed references to Dlite debugger

modified call to regs_init()

>>> memory.c

modified initial value of mem_stack_min

>>> regs.c

added extra dimension of MAX_PROCS to definitions of the following
 variables: regs_R, regs_F, regs_HI, regs_LO, regs_FCC, regs_PC

modified definition of regs_init() to accept three parameters:
  pid, initial stack pointer value, initial PC value (entry point)

modified body of regs_init() to use pid value
  for initializing correct registers

modified definition of regs_dump() to accept additional pid parameter
  and modified body to use this parameter to access appropriate registers

>>> stats.c

added new case for sc_uintarray in stat_eval_ident() and stat_delete()

added new function stat_reg_uintarray() to
 register an unsigned integer array statistical variable

in stat_print_stat(), added 'i' variable and case for sc_uintarray
 to dump array of integer values to output file

>>> syscall.c

added new typedef for nodes in queues associated with synchronization
 variables

added definitions for arrays of queues for synchronization variables
 (locks, barriers, semaphores)

added definition for array of flags indicating active processors

added definition for count of created processes

added function CreateNewProcess() that is invoked from ss_syscall()
 after simulated program calls run-time library function
   - allocates stack space for new thread of execution
   - initializes registers using regs_init()
   - performs other miscellaneous initialization tasks

modified ss_syscall() to accept additional parameter, the current pid
 for the thread of execution that invoked the system call

throughout the body of ss_syscall(), changed all references to regs_R
 array to include the pid parameter (*** NUMEROUS changes here ***)

in ss_syscall(), modified case for SS_SYS_exit to properly handle
 multiprocessor termination (i.e., waiting for all threads to finish)

in ss_syscall(), removed code for SS_SYS_fork, SS_SYS_vfork,
 and SS_SYS_sigstack, but kept the cases there (minor cleanup)

in ss_syscall(), added new cases for the following multiprocessor-related
 system calls:
   - SS_MP_ACQUIRE_LOCK
   - SS_MP_RELEASE_LOCK
   - SS_MP_INIT_LOCK
   - SS_MP_BARRIER
   - SS_MP_INIT_BARRIER
   - SS_MP_THREAD_ID
   - SS_MP_CREATE_THREAD
   - SS_MP_EXIT_THREAD
   - SS_MP_SEMA_WAIT
   - SS_MP_SEMA_SIGNAL
   - SS_MP_INIT_SEMA

>>> sysprobe.c

modified main() to return an integer value (this change was only made to
 eliminate compiler messages and has no bearing on multiprocessing)