watchpoint facility for ptx debugging

[git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 7160]

4 files changed, 252 insertions, 109 deletions

diff --git a/src/gpgpu-sim/Makefile b/src/gpgpu-sim/Makefile
index 2b1412a..96b1a5c 100644
--- a/src/gpgpu-sim/Makefile
+++ b/src/gpgpu-sim/Makefile
@@ -114,3 +114,251 @@ dram_sched.$(OEXT): ../cuda-sim/ptx.tab.h
 
 # DO NOT DELETE
 
+addrdec.o: /usr/include/string.h /usr/include/features.h
+addrdec.o: /usr/include/sys/cdefs.h /usr/include/bits/wordsize.h
+addrdec.o: /usr/include/gnu/stubs.h /usr/include/gnu/stubs-64.h addrdec.h
+addrdec.o: /usr/include/stdio.h /usr/include/bits/types.h
+addrdec.o: /usr/include/bits/typesizes.h /usr/include/libio.h
+addrdec.o: /usr/include/_G_config.h /usr/include/wchar.h
+addrdec.o: /usr/include/bits/stdio_lim.h /usr/include/bits/sys_errlist.h
+addrdec.o: /usr/include/stdlib.h /usr/include/sys/types.h /usr/include/time.h
+addrdec.o: /usr/include/endian.h /usr/include/bits/endian.h
+addrdec.o: /usr/include/bits/byteswap.h /usr/include/sys/select.h
+addrdec.o: /usr/include/bits/select.h /usr/include/bits/sigset.h
+addrdec.o: /usr/include/bits/time.h /usr/include/sys/sysmacros.h
+addrdec.o: /usr/include/bits/pthreadtypes.h /usr/include/alloca.h
+addrdec.o: /usr/include/assert.h ../option_parser.h
+delayqueue.o: delayqueue.h /usr/include/stdio.h /usr/include/features.h
+delayqueue.o: /usr/include/sys/cdefs.h /usr/include/bits/wordsize.h
+delayqueue.o: /usr/include/gnu/stubs.h /usr/include/gnu/stubs-64.h
+delayqueue.o: /usr/include/bits/types.h /usr/include/bits/typesizes.h
+delayqueue.o: /usr/include/libio.h /usr/include/_G_config.h
+delayqueue.o: /usr/include/wchar.h /usr/include/bits/stdio_lim.h
+delayqueue.o: /usr/include/bits/sys_errlist.h /usr/include/assert.h
+delayqueue.o: /usr/include/stdlib.h /usr/include/sys/types.h
+delayqueue.o: /usr/include/time.h /usr/include/endian.h
+delayqueue.o: /usr/include/bits/endian.h /usr/include/bits/byteswap.h
+delayqueue.o: /usr/include/sys/select.h /usr/include/bits/select.h
+delayqueue.o: /usr/include/bits/sigset.h /usr/include/bits/time.h
+delayqueue.o: /usr/include/sys/sysmacros.h /usr/include/bits/pthreadtypes.h
+delayqueue.o: /usr/include/alloca.h gpu-misc.h ../intersim/statwraper.h
+dram.o: gpu-misc.h dram.h /usr/include/stdio.h /usr/include/features.h
+dram.o: /usr/include/sys/cdefs.h /usr/include/bits/wordsize.h
+dram.o: /usr/include/gnu/stubs.h /usr/include/gnu/stubs-64.h
+dram.o: /usr/include/bits/types.h /usr/include/bits/typesizes.h
+dram.o: /usr/include/libio.h /usr/include/_G_config.h /usr/include/wchar.h
+dram.o: /usr/include/bits/stdio_lim.h /usr/include/bits/sys_errlist.h
+dram.o: /usr/include/stdlib.h /usr/include/sys/types.h /usr/include/time.h
+dram.o: /usr/include/endian.h /usr/include/bits/endian.h
+dram.o: /usr/include/bits/byteswap.h /usr/include/sys/select.h
+dram.o: /usr/include/bits/select.h /usr/include/bits/sigset.h
+dram.o: /usr/include/bits/time.h /usr/include/sys/sysmacros.h
+dram.o: /usr/include/bits/pthreadtypes.h /usr/include/alloca.h delayqueue.h
+dram.o: /usr/include/assert.h ../cuda-sim/dram_callback.h
+dram_sched.o: dram_sched.h dram.h /usr/include/stdio.h
+dram_sched.o: /usr/include/features.h /usr/include/sys/cdefs.h
+dram_sched.o: /usr/include/bits/wordsize.h /usr/include/gnu/stubs.h
+dram_sched.o: /usr/include/gnu/stubs-64.h /usr/include/bits/types.h
+dram_sched.o: /usr/include/bits/typesizes.h /usr/include/libio.h
+dram_sched.o: /usr/include/_G_config.h /usr/include/wchar.h
+dram_sched.o: /usr/include/bits/stdio_lim.h /usr/include/bits/sys_errlist.h
+dram_sched.o: /usr/include/stdlib.h /usr/include/sys/types.h
+dram_sched.o: /usr/include/time.h /usr/include/endian.h
+dram_sched.o: /usr/include/bits/endian.h /usr/include/bits/byteswap.h
+dram_sched.o: /usr/include/sys/select.h /usr/include/bits/select.h
+dram_sched.o: /usr/include/bits/sigset.h /usr/include/bits/time.h
+dram_sched.o: /usr/include/sys/sysmacros.h /usr/include/bits/pthreadtypes.h
+dram_sched.o: /usr/include/alloca.h delayqueue.h /usr/include/assert.h
+dram_sched.o: ../cuda-sim/dram_callback.h shader.h /usr/include/math.h
+dram_sched.o: /usr/include/bits/huge_val.h /usr/include/bits/mathdef.h
+dram_sched.o: /usr/include/bits/mathcalls.h /usr/include/limits.h
+dram_sched.o: /usr/include/bits/posix1_lim.h /usr/include/bits/local_lim.h
+dram_sched.o: /usr/include/linux/limits.h /usr/include/bits/posix2_lim.h
+dram_sched.o: ../cuda-sim/ptx.tab.h gpu-cache.h ../abstract_hardware_model.h
+dram_sched.o: stack.h gpu-sim.h /usr/include/zlib.h /usr/include/zconf.h
+dram_sched.o: /usr/include/unistd.h /usr/include/bits/posix_opt.h
+dram_sched.o: /usr/include/bits/confname.h /usr/include/getopt.h
+dram_sched.o: ../option_parser.h mem_fetch.h addrdec.h gpu-misc.h
+dwf.o: dwf.h ../abstract_hardware_model.h histogram.h /usr/include/string.h
+dwf.o: /usr/include/features.h /usr/include/sys/cdefs.h
+dwf.o: /usr/include/bits/wordsize.h /usr/include/gnu/stubs.h
+dwf.o: /usr/include/gnu/stubs-64.h
+gpu-cache.o: gpu-cache.h /usr/include/stdio.h /usr/include/features.h
+gpu-cache.o: /usr/include/sys/cdefs.h /usr/include/bits/wordsize.h
+gpu-cache.o: /usr/include/gnu/stubs.h /usr/include/gnu/stubs-64.h
+gpu-cache.o: /usr/include/bits/types.h /usr/include/bits/typesizes.h
+gpu-cache.o: /usr/include/libio.h /usr/include/_G_config.h
+gpu-cache.o: /usr/include/wchar.h /usr/include/bits/stdio_lim.h
+gpu-cache.o: /usr/include/bits/sys_errlist.h /usr/include/stdlib.h
+gpu-cache.o: /usr/include/sys/types.h /usr/include/time.h
+gpu-cache.o: /usr/include/endian.h /usr/include/bits/endian.h
+gpu-cache.o: /usr/include/bits/byteswap.h /usr/include/sys/select.h
+gpu-cache.o: /usr/include/bits/select.h /usr/include/bits/sigset.h
+gpu-cache.o: /usr/include/bits/time.h /usr/include/sys/sysmacros.h
+gpu-cache.o: /usr/include/bits/pthreadtypes.h /usr/include/alloca.h
+gpu-cache.o: ../abstract_hardware_model.h gpu-misc.h addrdec.h
+gpu-cache.o: /usr/include/assert.h ../option_parser.h /usr/include/string.h
+gpu-misc.o: gpu-misc.h
+gpu-sim.o: gpu-sim.h /usr/include/stdio.h /usr/include/features.h
+gpu-sim.o: /usr/include/sys/cdefs.h /usr/include/bits/wordsize.h
+gpu-sim.o: /usr/include/gnu/stubs.h /usr/include/gnu/stubs-64.h
+gpu-sim.o: /usr/include/bits/types.h /usr/include/bits/typesizes.h
+gpu-sim.o: /usr/include/libio.h /usr/include/_G_config.h /usr/include/wchar.h
+gpu-sim.o: /usr/include/bits/stdio_lim.h /usr/include/bits/sys_errlist.h
+gpu-sim.o: /usr/include/stdlib.h /usr/include/sys/types.h /usr/include/time.h
+gpu-sim.o: /usr/include/endian.h /usr/include/bits/endian.h
+gpu-sim.o: /usr/include/bits/byteswap.h /usr/include/sys/select.h
+gpu-sim.o: /usr/include/bits/select.h /usr/include/bits/sigset.h
+gpu-sim.o: /usr/include/bits/time.h /usr/include/sys/sysmacros.h
+gpu-sim.o: /usr/include/bits/pthreadtypes.h /usr/include/alloca.h
+gpu-sim.o: /usr/include/math.h /usr/include/bits/huge_val.h
+gpu-sim.o: /usr/include/bits/mathdef.h /usr/include/bits/mathcalls.h
+gpu-sim.o: /usr/include/zlib.h /usr/include/zconf.h /usr/include/unistd.h
+gpu-sim.o: /usr/include/bits/posix_opt.h /usr/include/bits/confname.h
+gpu-sim.o: /usr/include/getopt.h ../option_parser.h shader.h
+gpu-sim.o: /usr/include/limits.h /usr/include/bits/posix1_lim.h
+gpu-sim.o: /usr/include/bits/local_lim.h /usr/include/linux/limits.h
+gpu-sim.o: /usr/include/bits/posix2_lim.h /usr/include/assert.h
+gpu-sim.o: ../cuda-sim/ptx.tab.h ../cuda-sim/dram_callback.h gpu-cache.h
+gpu-sim.o: ../abstract_hardware_model.h delayqueue.h stack.h dram.h
+gpu-sim.o: mem_fetch.h addrdec.h gpu-misc.h icnt_wrapper.h dwf.h
+gpu-sim.o: warp_tracker.h cflogger.h l2cache.h ../cuda-sim/ptx-stats.h
+gpu-sim.o: ../intersim/statwraper.h ../debug.h ../abstract_hardware_model.h
+gpu-sim.o: /usr/include/string.h mem_latency_stat.h
+icnt_wrapper.o: icnt_wrapper.h /usr/include/assert.h /usr/include/features.h
+icnt_wrapper.o: /usr/include/sys/cdefs.h /usr/include/bits/wordsize.h
+icnt_wrapper.o: /usr/include/gnu/stubs.h /usr/include/gnu/stubs-64.h
+icnt_wrapper.o: ../intersim/interconnect_interface.h ../intersim/flit.hpp
+icnt_wrapper.o: ../intersim/booksim.hpp ../intersim/trafficmanager.hpp
+icnt_wrapper.o: ../intersim/module.hpp ../intersim/config_utils.hpp
+icnt_wrapper.o: /usr/include/stdio.h /usr/include/bits/types.h
+icnt_wrapper.o: /usr/include/bits/typesizes.h /usr/include/libio.h
+icnt_wrapper.o: /usr/include/_G_config.h /usr/include/wchar.h
+icnt_wrapper.o: /usr/include/bits/stdio_lim.h /usr/include/bits/sys_errlist.h
+icnt_wrapper.o: ../intersim/network.hpp ../intersim/credit.hpp
+icnt_wrapper.o: ../intersim/router.hpp ../intersim/buffer_state.hpp
+icnt_wrapper.o: ../intersim/stats.hpp ../intersim/traffic.hpp
+icnt_wrapper.o: ../intersim/routefunc.hpp ../intersim/outputset.hpp
+icnt_wrapper.o: ../intersim/injection.hpp
+l2cache.o: /usr/include/stdlib.h /usr/include/features.h
+l2cache.o: /usr/include/sys/cdefs.h /usr/include/bits/wordsize.h
+l2cache.o: /usr/include/gnu/stubs.h /usr/include/gnu/stubs-64.h
+l2cache.o: /usr/include/sys/types.h /usr/include/bits/types.h
+l2cache.o: /usr/include/bits/typesizes.h /usr/include/time.h
+l2cache.o: /usr/include/endian.h /usr/include/bits/endian.h
+l2cache.o: /usr/include/bits/byteswap.h /usr/include/sys/select.h
+l2cache.o: /usr/include/bits/select.h /usr/include/bits/sigset.h
+l2cache.o: /usr/include/bits/time.h /usr/include/sys/sysmacros.h
+l2cache.o: /usr/include/bits/pthreadtypes.h /usr/include/alloca.h
+l2cache.o: /usr/include/stdio.h /usr/include/libio.h /usr/include/_G_config.h
+l2cache.o: /usr/include/wchar.h /usr/include/bits/stdio_lim.h
+l2cache.o: /usr/include/bits/sys_errlist.h /usr/include/string.h
+l2cache.o: ../tr1_hash_map.h ../option_parser.h mem_fetch.h shader.h
+l2cache.o: /usr/include/math.h /usr/include/bits/huge_val.h
+l2cache.o: /usr/include/bits/mathdef.h /usr/include/bits/mathcalls.h
+l2cache.o: /usr/include/limits.h /usr/include/bits/posix1_lim.h
+l2cache.o: /usr/include/bits/local_lim.h /usr/include/linux/limits.h
+l2cache.o: /usr/include/bits/posix2_lim.h /usr/include/assert.h
+l2cache.o: ../cuda-sim/ptx.tab.h ../cuda-sim/dram_callback.h gpu-cache.h
+l2cache.o: ../abstract_hardware_model.h delayqueue.h stack.h dram.h addrdec.h
+l2cache.o: histogram.h l2cache.h ../intersim/statwraper.h
+shader.o: shader.h /usr/include/stdio.h /usr/include/features.h
+shader.o: /usr/include/sys/cdefs.h /usr/include/bits/wordsize.h
+shader.o: /usr/include/gnu/stubs.h /usr/include/gnu/stubs-64.h
+shader.o: /usr/include/bits/types.h /usr/include/bits/typesizes.h
+shader.o: /usr/include/libio.h /usr/include/_G_config.h /usr/include/wchar.h
+shader.o: /usr/include/bits/stdio_lim.h /usr/include/bits/sys_errlist.h
+shader.o: /usr/include/stdlib.h /usr/include/sys/types.h /usr/include/time.h
+shader.o: /usr/include/endian.h /usr/include/bits/endian.h
+shader.o: /usr/include/bits/byteswap.h /usr/include/sys/select.h
+shader.o: /usr/include/bits/select.h /usr/include/bits/sigset.h
+shader.o: /usr/include/bits/time.h /usr/include/sys/sysmacros.h
+shader.o: /usr/include/bits/pthreadtypes.h /usr/include/alloca.h
+shader.o: /usr/include/math.h /usr/include/bits/huge_val.h
+shader.o: /usr/include/bits/mathdef.h /usr/include/bits/mathcalls.h
+shader.o: /usr/include/limits.h /usr/include/bits/posix1_lim.h
+shader.o: /usr/include/bits/local_lim.h /usr/include/linux/limits.h
+shader.o: /usr/include/bits/posix2_lim.h /usr/include/assert.h
+shader.o: ../cuda-sim/ptx.tab.h ../cuda-sim/dram_callback.h gpu-cache.h
+shader.o: ../abstract_hardware_model.h delayqueue.h stack.h dram.h gpu-sim.h
+shader.o: /usr/include/zlib.h /usr/include/zconf.h /usr/include/unistd.h
+shader.o: /usr/include/bits/posix_opt.h /usr/include/bits/confname.h
+shader.o: /usr/include/getopt.h ../option_parser.h mem_fetch.h addrdec.h
+shader.o: dwf.h warp_tracker.h cflogger.h gpu-misc.h ../cuda-sim/ptx_sim.h
+shader.o: ../cuda-sim/dram_callback.h ../cuda-sim/opcodes.h
+shader.o: ../cuda-sim/opcodes.def ../cuda-sim/ptx-stats.h
+shader.o: /usr/include/string.h
+stack.o: stack.h ../abstract_hardware_model.h /usr/include/stdlib.h
+stack.o: /usr/include/features.h /usr/include/sys/cdefs.h
+stack.o: /usr/include/bits/wordsize.h /usr/include/gnu/stubs.h
+stack.o: /usr/include/gnu/stubs-64.h /usr/include/sys/types.h
+stack.o: /usr/include/bits/types.h /usr/include/bits/typesizes.h
+stack.o: /usr/include/time.h /usr/include/endian.h /usr/include/bits/endian.h
+stack.o: /usr/include/bits/byteswap.h /usr/include/sys/select.h
+stack.o: /usr/include/bits/select.h /usr/include/bits/sigset.h
+stack.o: /usr/include/bits/time.h /usr/include/sys/sysmacros.h
+stack.o: /usr/include/bits/pthreadtypes.h /usr/include/alloca.h
+stack.o: /usr/include/assert.h
+stat-tool.o: /usr/include/stdio.h /usr/include/features.h
+stat-tool.o: /usr/include/sys/cdefs.h /usr/include/bits/wordsize.h
+stat-tool.o: /usr/include/gnu/stubs.h /usr/include/gnu/stubs-64.h
+stat-tool.o: /usr/include/bits/types.h /usr/include/bits/typesizes.h
+stat-tool.o: /usr/include/libio.h /usr/include/_G_config.h
+stat-tool.o: /usr/include/wchar.h /usr/include/bits/stdio_lim.h
+stat-tool.o: /usr/include/bits/sys_errlist.h /usr/include/stdlib.h
+stat-tool.o: /usr/include/sys/types.h /usr/include/time.h
+stat-tool.o: /usr/include/endian.h /usr/include/bits/endian.h
+stat-tool.o: /usr/include/bits/byteswap.h /usr/include/sys/select.h
+stat-tool.o: /usr/include/bits/select.h /usr/include/bits/sigset.h
+stat-tool.o: /usr/include/bits/time.h /usr/include/sys/sysmacros.h
+stat-tool.o: /usr/include/bits/pthreadtypes.h /usr/include/alloca.h
+stat-tool.o: /usr/include/assert.h /usr/include/zlib.h /usr/include/zconf.h
+stat-tool.o: /usr/include/unistd.h /usr/include/bits/posix_opt.h
+stat-tool.o: /usr/include/bits/confname.h /usr/include/getopt.h histogram.h
+stat-tool.o: ../abstract_hardware_model.h cflogger.h
+visualizer.o: gpu-sim.h /usr/include/stdio.h /usr/include/features.h
+visualizer.o: /usr/include/sys/cdefs.h /usr/include/bits/wordsize.h
+visualizer.o: /usr/include/gnu/stubs.h /usr/include/gnu/stubs-64.h
+visualizer.o: /usr/include/bits/types.h /usr/include/bits/typesizes.h
+visualizer.o: /usr/include/libio.h /usr/include/_G_config.h
+visualizer.o: /usr/include/wchar.h /usr/include/bits/stdio_lim.h
+visualizer.o: /usr/include/bits/sys_errlist.h /usr/include/stdlib.h
+visualizer.o: /usr/include/sys/types.h /usr/include/time.h
+visualizer.o: /usr/include/endian.h /usr/include/bits/endian.h
+visualizer.o: /usr/include/bits/byteswap.h /usr/include/sys/select.h
+visualizer.o: /usr/include/bits/select.h /usr/include/bits/sigset.h
+visualizer.o: /usr/include/bits/time.h /usr/include/sys/sysmacros.h
+visualizer.o: /usr/include/bits/pthreadtypes.h /usr/include/alloca.h
+visualizer.o: /usr/include/math.h /usr/include/bits/huge_val.h
+visualizer.o: /usr/include/bits/mathdef.h /usr/include/bits/mathcalls.h
+visualizer.o: /usr/include/zlib.h /usr/include/zconf.h /usr/include/unistd.h
+visualizer.o: /usr/include/bits/posix_opt.h /usr/include/bits/confname.h
+visualizer.o: /usr/include/getopt.h ../option_parser.h shader.h
+visualizer.o: /usr/include/limits.h /usr/include/bits/posix1_lim.h
+visualizer.o: /usr/include/bits/local_lim.h /usr/include/linux/limits.h
+visualizer.o: /usr/include/bits/posix2_lim.h /usr/include/assert.h
+visualizer.o: ../cuda-sim/ptx.tab.h ../cuda-sim/dram_callback.h gpu-cache.h
+visualizer.o: ../abstract_hardware_model.h delayqueue.h stack.h dram.h
+visualizer.o: mem_fetch.h addrdec.h /usr/include/string.h
+visualizer.o: ../gpgpu-sim/shader.h ../gpgpu-sim/mem_fetch.h
+warp_tracker.o: warp_tracker.h ../abstract_hardware_model.h shader.h
+warp_tracker.o: /usr/include/stdio.h /usr/include/features.h
+warp_tracker.o: /usr/include/sys/cdefs.h /usr/include/bits/wordsize.h
+warp_tracker.o: /usr/include/gnu/stubs.h /usr/include/gnu/stubs-64.h
+warp_tracker.o: /usr/include/bits/types.h /usr/include/bits/typesizes.h
+warp_tracker.o: /usr/include/libio.h /usr/include/_G_config.h
+warp_tracker.o: /usr/include/wchar.h /usr/include/bits/stdio_lim.h
+warp_tracker.o: /usr/include/bits/sys_errlist.h /usr/include/stdlib.h
+warp_tracker.o: /usr/include/sys/types.h /usr/include/time.h
+warp_tracker.o: /usr/include/endian.h /usr/include/bits/endian.h
+warp_tracker.o: /usr/include/bits/byteswap.h /usr/include/sys/select.h
+warp_tracker.o: /usr/include/bits/select.h /usr/include/bits/sigset.h
+warp_tracker.o: /usr/include/bits/time.h /usr/include/sys/sysmacros.h
+warp_tracker.o: /usr/include/bits/pthreadtypes.h /usr/include/alloca.h
+warp_tracker.o: /usr/include/math.h /usr/include/bits/huge_val.h
+warp_tracker.o: /usr/include/bits/mathdef.h /usr/include/bits/mathcalls.h
+warp_tracker.o: /usr/include/limits.h /usr/include/bits/posix1_lim.h
+warp_tracker.o: /usr/include/bits/local_lim.h /usr/include/linux/limits.h
+warp_tracker.o: /usr/include/bits/posix2_lim.h /usr/include/assert.h
+warp_tracker.o: ../cuda-sim/ptx.tab.h ../cuda-sim/dram_callback.h gpu-cache.h
+warp_tracker.o: delayqueue.h stack.h dram.h
diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc
index 2b4ebe3..1bec72f 100644
--- a/src/gpgpu-sim/gpu-sim.cc
+++ b/src/gpgpu-sim/gpu-sim.cc
@@ -82,6 +82,7 @@
 #include "../cuda-sim/ptx-stats.h"
 #include "../intersim/statwraper.h"
 #include "../abstract_hardware_model.h"
+#include "../debug.h"
 
 #include <stdio.h>
 #include <string.h>
@@ -206,7 +207,6 @@ void check_time_vector_update(unsigned int uid,int slot ,long int latency,int ty
 void node_req_hist_clear(void *p);
 void node_req_hist_dump(void *p);
 void node_req_hist_update(void * p,int node, long long cycle);
-static void gpgpu_debug();
 
 /* functionally simulated memory */
 extern struct mem_t *mem;
@@ -1765,111 +1765,3 @@ void dump_pipeline()
    dump_pipeline_impl(0,-1,-1);
 }
 
-/// interactive debugger 
-
-static void gpgpu_debug()
-{
-   bool done=true;
-
-   static bool single_step=true;
-   static unsigned next_brkpt=1;
-   static std::map<unsigned,brk_pt> breakpoints;
-
-   /// if single stepping, go to interactive debugger
-
-   if( single_step ) 
-      done=false;
-
-   /// check if we've reached a breakpoint
-
-   for( std::map<unsigned,brk_pt>::iterator i=breakpoints.begin(); i!=breakpoints.end(); i++) {
-      unsigned num=i->first;
-      brk_pt &b=i->second;
-      if( b.is_watchpoint() ) {
-         unsigned addr = b.get_addr();
-         unsigned new_value = read_location(addr);
-         if( new_value != b.get_value() ) {
-            printf( "GPGPU-Sim DBG: watch point %u triggered (old value=%x, new value=%x)\n",
-                     num,b.get_value(),new_value );
-            b.set_value(new_value);
-            done = false; 
-         }
-      } else {
-         for( unsigned sid=0; sid < gpu_n_shader; sid++ ) { 
-            inst_t *fvi = first_valid_thread(sc[sid]->pipeline_reg[IF_ID]);
-            if( !fvi ) continue;
-            if( thread_at_brkpt(fvi->ptx_thd_info, b) ) {
-               done = false;
-               printf("GPGPU-Sim DBG: reached breakpoint %u at %s (sm=%u, hwtid=%u)\n", 
-                      num, b.location().c_str(), sid, fvi->hw_thread_id );
-            }
-         }
-      }
-   }
-
-   /// enter interactive debugger loop
-
-   while (!done) {
-      printf("(gpgpu-sim dbg) ");
-      fflush(stdout);
-      
-      char line[1024];
-      fgets(line,1024,stdin);
-
-      char *tok = strtok(line," \t\n");
-      if( !strcmp(tok,"dp") ) {
-         int shader_num = 0;
-         tok = strtok(NULL," \t\n");
-         sscanf(tok,"%d",&shader_num);
-         dump_pipeline_impl((0x40|0x4|0x1),shader_num,0);
-         printf("\n");
-         fflush(stdout);
-      } else if( !strcmp(tok,"q") || !strcmp(tok,"quit") ) {
-         printf("\nreally quit GPGPU-Sim (y/n)?\n");
-         fgets(line,1024,stdin);
-         tok = strtok(NULL," \t\n");
-         if( !strcmp(tok,"y") ) {
-            exit(0);
-         } else {
-            printf("not quiting.\n");
-         }
-      } else if( !strcmp(tok,"b") ) {
-         tok = strtok(NULL," \t\n");
-         char brkpt[1024];
-         sscanf(tok,"%s",brkpt);
-         tok = strtok(NULL," \t\n");
-         unsigned uid;
-         sscanf(tok,"%u",&uid);
-         breakpoints[next_brkpt++] = brk_pt(brkpt,uid);
-      } else if( !strcmp(tok,"d") ) {
-         tok = strtok(NULL," \t\n");
-         unsigned uid;
-         sscanf(tok,"%u",&uid);
-         breakpoints.erase(uid);
-      } else if( !strcmp(tok,"s") ) {
-         done = true;
-      } else if( !strcmp(tok,"c") ) {
-         single_step=false;
-         done = true;
-      } else if( !strcmp(tok,"w") ) {
-         tok = strtok(NULL," \t\n");
-         unsigned addr;
-         sscanf(tok,"%x",&addr);
-         unsigned value = read_location(addr);
-         breakpoints[next_brkpt++] = brk_pt(addr,value);
-      } else if( !strcmp(tok,"h") ) {
-         printf("commands:\n");
-         printf("  q                           - quit GPGPU-Sim\n");
-         printf("  b <file>:<line> <thead uid> - set breakpoint\n");
-         printf("  w <global address>          - set watchpoint\n");
-         printf("  del <n>                     - delete breakpoint\n");
-         printf("  s                           - single step one shader cycle (all cores)\n");
-         printf("  c                           - continue simulation without single stepping\n");
-         printf("  dp <n>                      - display pipeline contents on SM <n>\n");
-         printf("  h                           - print this message\n");
-      } else {
-         printf("\ncommand not understood.\n");
-      }
-      fflush(stdout);
-   }
-}
diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h
index e30797e..27d83ad 100644
--- a/src/gpgpu-sim/gpu-sim.h
+++ b/src/gpgpu-sim/gpu-sim.h
@@ -112,5 +112,6 @@ void gpu_print_stat();
 int mem_ctrl_full( int mc_id );
 
 void dramqueue_latency_log_dump();
+void dump_pipeline_impl( int mask, int s, int m );
 
 #endif
diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h
index 9f64da5..9dd7dc8 100644
--- a/src/gpgpu-sim/shader.h
+++ b/src/gpgpu-sim/shader.h
@@ -896,5 +896,7 @@ unsigned int max_cta_per_shader( shader_core_ctx_t *shader);
 #define WB_RT 6
 #define ID_OC 7
 
+extern unsigned int gpu_n_shader;
+extern shader_core_ctx_t **sc;
 
 #endif /* SHADER_H */