added a first attempt at code prefetching

Author: sethpugsley

build_champsim.sh

@@ -1,18 +1,19 @@
 #!/bin/bash
 
-if [ "$#" -ne 6 ]; then
+if [ "$#" -ne 7 ]; then
     echo "Illegal number of parameters"
     echo "Usage: ./build_champsim.sh [branch_pred] [l1d_pref] [l2c_pref] [llc_pref] [llc_repl] [num_core]"
     exit 1
 fi
 
 # ChampSim configuration
 BRANCH=$1           # branch/*.bpred
-L1D_PREFETCHER=$2   # prefetcher/*.l1d_pref
-L2C_PREFETCHER=$3   # prefetcher/*.l2c_pref
-LLC_PREFETCHER=$4   # prefetcher/*.llc_pref
-LLC_REPLACEMENT=$5  # replacement/*.llc_repl
-NUM_CORE=$6         # tested up to 8-core system
+L1I_PREFETCHER=$2   # prefetcher/*.l1i_pref
+L1D_PREFETCHER=$3   # prefetcher/*.l1d_pref
+L2C_PREFETCHER=$4   # prefetcher/*.l2c_pref
+LLC_PREFETCHER=$5   # prefetcher/*.llc_pref
+LLC_REPLACEMENT=$6  # replacement/*.llc_repl
+NUM_CORE=$7         # tested up to 8-core system
 
 ############## Some useful macros ###############
 BOLD=$(tput bold)
@@ -27,6 +28,13 @@ if [ ! -f ./branch/${BRANCH}.bpred ]; then
     exit 1
 fi
 
+if [ ! -f ./prefetcher/${L1I_PREFETCHER}.l1i_pref ]; then
+    echo "[ERROR] Cannot find L1I prefetcher"
+	echo "[ERROR] Possible L1I prefetchers from prefetcher/*.l1i_pref "
+    find prefetcher -name "*.l1i_pref"
+    exit 1
+fi
+
 if [ ! -f ./prefetcher/${L1D_PREFETCHER}.l1d_pref ]; then
     echo "[ERROR] Cannot find L1D prefetcher"
 	echo "[ERROR] Possible L1D prefetchers from prefetcher/*.l1d_pref "
@@ -80,6 +88,7 @@ echo
 
 # Change prefetchers and replacement policy
 cp branch/${BRANCH}.bpred branch/branch_predictor.cc
+cp prefetcher/${L1I_PREFETCHER}.l1i_pref prefetcher/l1i_prefetcher.cc
 cp prefetcher/${L1D_PREFETCHER}.l1d_pref prefetcher/l1d_prefetcher.cc
 cp prefetcher/${L2C_PREFETCHER}.l2c_pref prefetcher/l2c_prefetcher.cc
 cp prefetcher/${LLC_PREFETCHER}.llc_pref prefetcher/llc_prefetcher.cc
@@ -101,12 +110,13 @@ fi
 
 echo "${BOLD}ChampSim is successfully built"
 echo "Branch Predictor: ${BRANCH}"
+echo "L1I Prefetcher: ${L1I_PREFETCHER}"
 echo "L1D Prefetcher: ${L1D_PREFETCHER}"
 echo "L2C Prefetcher: ${L2C_PREFETCHER}"
 echo "LLC Prefetcher: ${LLC_PREFETCHER}"
 echo "LLC Replacement: ${LLC_REPLACEMENT}"
 echo "Cores: ${NUM_CORE}"
-BINARY_NAME="${BRANCH}-${L1D_PREFETCHER}-${L2C_PREFETCHER}-${LLC_PREFETCHER}-${LLC_REPLACEMENT}-${NUM_CORE}core"
+BINARY_NAME="${BRANCH}-${L1I_PREFETCHER}-${L1D_PREFETCHER}-${L2C_PREFETCHER}-${LLC_PREFETCHER}-${LLC_REPLACEMENT}-${NUM_CORE}core"
 echo "Binary: bin/${BINARY_NAME}"
 echo ""
 mv bin/champsim bin/${BINARY_NAME}
@@ -118,6 +128,7 @@ sed -i.bak 's/\<NUM_CPUS '${NUM_CORE}'\>/NUM_CPUS 1/g' inc/champsim.h
 #sed -i.bak 's/\<DRAM_CHANNELS_LOG2 1\>/DRAM_CHANNELS_LOG2 0/g' inc/champsim.h
 
 cp branch/bimodal.bpred branch/branch_predictor.cc
+cp prefetcher/no.l1i_pref prefetcher/l1i_prefetcher.cc
 cp prefetcher/no.l1d_pref prefetcher/l1d_prefetcher.cc
 cp prefetcher/no.l2c_pref prefetcher/l2c_prefetcher.cc
 cp prefetcher/no.llc_pref prefetcher/llc_prefetcher.cc

inc/ooo_cpu.h

@@ -23,6 +23,7 @@ using namespace std;
 #define BRANCH_MISPREDICT_PENALTY 1
 #define DECODED_INSTRUCTION_CACHE_SIZE 2048
 #define IFETCH_WORKING_CACHE_LINES 4
+#define CODE_PREFETCH_BUFFER_SIZE 16
 //#define SCHEDULING_LATENCY 0
 //#define EXEC_LATENCY 0
 //#define DECODE_LATENCY 2
@@ -63,7 +64,11 @@ class O3_CPU {
   // current working cache lines for instruction fetch stage
   ooo_model_instr ifetch_working_set_instrs[IFETCH_WORKING_CACHE_LINES];
   uint32_t ifetch_working_set_mru_index;
-  
+
+  // code prefetching
+  ooo_model_instr code_prefetch_buffer[CODE_PREFETCH_BUFFER_SIZE];
+  uint32_t code_prefetch_buffer_occupancy;  
+
     // store array, this structure is required to properly handle store instructions
     uint64_t STA[STA_SIZE], STA_head, STA_tail; 
 
@@ -186,6 +191,15 @@ class O3_CPU {
 	    ifetch_working_set_instrs[i].fetched = COMPLETED;
 	  }
 	ifetch_working_set_mru_index = 0;
+
+	for(int i=0; i<CODE_PREFETCH_BUFFER_SIZE; i++)
+	  {
+	    code_prefetch_buffer[i].ip = 0;
+	    code_prefetch_buffer[i].branch_target = 0;
+	    code_prefetch_buffer[i].translated = 0;
+	    code_prefetch_buffer[i].fetched = 0;
+	  }
+	code_prefetch_buffer_occupancy = 0;
     }
 
     // functions
@@ -233,7 +247,14 @@ class O3_CPU {
     // branch predictor
     uint8_t predict_branch(uint64_t ip);
     void    initialize_branch_predictor(),
-            last_branch_result(uint64_t ip, uint8_t taken); 
+            last_branch_result(uint64_t ip, uint8_t taken);
+
+  // code prefetching
+  void l1i_prefetcher_initialize();
+  void l1i_prefetcher_branch_operate(uint64_t ip, uint8_t branch_type, uint8_t branch_prediction);
+  void l1i_prefetcher_cache_operate(uint64_t addr, uint8_t cache_hit);
+  void l1i_prefetcher_final_stats();
+  int prefetch_code_line(uint64_t ip, uint64_t pf_addr); 
 };
 
 extern O3_CPU ooo_cpu[NUM_CPUS];

prefetcher/l1i_prefetcher.cc

@@ -0,0 +1,21 @@
+#include "ooo_cpu.h"
+
+void O3_CPU::l1i_prefetcher_initialize() 
+{
+
+}
+
+void O3_CPU::l1i_prefetcher_branch_operate(uint64_t ip, uint8_t branch_type, uint8_t branch_prediction)
+{
+
+}
+
+void O3_CPU::l1i_prefetcher_cache_operate(uint64_t addr, uint8_t cache_hit)
+{
+
+}
+
+void O3_CPU::l1i_prefetcher_final_stats()
+{
+
+}

prefetcher/next_line.l1i_pref

@@ -0,0 +1,25 @@
+#include "ooo_cpu.h"
+
+void O3_CPU::l1i_prefetcher_initialize() 
+{
+
+}
+
+void O3_CPU::l1i_prefetcher_branch_operate(uint64_t ip, uint8_t branch_type, uint8_t branch_prediction)
+{
+
+}
+
+void O3_CPU::l1i_prefetcher_cache_operate(uint64_t addr, uint8_t cache_hit)
+{
+  if((cache_hit == 0) && (L1I.MSHR.occupancy < (L1I.MSHR.SIZE>>1)))
+    {
+      uint64_t pf_addr = addr + (1<<LOG2_BLOCK_SIZE);
+      prefetch_code_line(addr, pf_addr);
+    }
+}
+
+void O3_CPU::l1i_prefetcher_final_stats()
+{
+
+}

prefetcher/no.l1i_pref

@@ -0,0 +1,21 @@
+#include "ooo_cpu.h"
+
+void O3_CPU::l1i_prefetcher_initialize() 
+{
+
+}
+
+void O3_CPU::l1i_prefetcher_branch_operate(uint64_t ip, uint8_t branch_type, uint8_t branch_prediction)
+{
+
+}
+
+void O3_CPU::l1i_prefetcher_cache_operate(uint64_t addr, uint8_t cache_hit)
+{
+
+}
+
+void O3_CPU::l1i_prefetcher_final_stats()
+{
+
+}