Remove unused variables

8086_assembly_01/Makefile

@@ -1,2 +1,2 @@
 all:
-	clang++ -g dasm.cpp -o dasm
+	clang++ -g dasm.cc -o dasm

8086_assembly_02/Makefile

@@ -1,2 +1,2 @@
 all:
-	clang++ -g dasm.cpp -o dasm
+	clang++ -g dasm.cc -o dasm

8086_assembly_03/Makefile

@@ -1,2 +1,2 @@
 all:
-	clang++ -g dasm.cpp -o dasm
+	clang++ -g dasm.cc -o dasm

8086_sim/Makefile

@@ -1,7 +1,7 @@
 CC=clang++
 CFLAGS=-g -O0 -Wall -Wextra
 LIBS=-Wl,-rpath,./lib -L./lib -lsim86
-SRC=*.cpp
+SRC=*.cc
 OUT=sim86
 
 all:

haversine_01/Makefile

@@ -1,17 +1,17 @@
 mk_haversine_fscanf:
-	clang++ -g cpp/fscanf.cpp cpp/haversine.cpp -o cpp/haverscan
+	clang++ -g cpp/fscanf.cc cpp/haversine.cc -o cpp/haverscan
 
 run_haversine_fscanf:
 	cd ./cpp && ./haverscan
 
 mk_haversine_strtok:
-	clang++ -g cpp/strtok.cpp cpp/haversine.cpp -o cpp/haverstrtok
+	clang++ -g cpp/strtok.cc cpp/haversine.cc -o cpp/haverstrtok
 
 run_haversine_strtok:
 	cd ./cpp && ./haverstrtok
 
 mk_test:
-	clang++ -g -lpthread cpp/test.cpp cpp/haversine.cpp -o cpp/test
+	clang++ -g -lpthread cpp/test.cc cpp/haversine.cc -o cpp/test
 
 run_test:
 	cd ./cpp && ./test

haversine_02/.gitignore

@@ -4,6 +4,7 @@
 compile_commands.json
 count_and_distances
 pairs.json
+cache_test
 main
 genhavr
 prochavr

haversine_02/compile

@@ -43,11 +43,11 @@ fi
 
 
 # GENERATOR
-GENSRC="./src/generator/gen_argparser.cpp \
-	./src/generator/generator.cpp \
-	./src/haversine.cpp \
-	./src/point_types.cpp \
-	./src/generator/main.cpp"
+GENSRC="./src/generator/gen_argparser.cc \
+	./src/generator/generator.cc \
+	./src/haversine.cc \
+	./src/point_types.cc \
+	./src/generator/main.cc"
 GENOUT=genhavr
 
 (set -x ; $CXX $CFLAGS $GENSRC -o $GENOUT)
@@ -64,10 +64,10 @@ JSONFLAGS="-c "
 JSON_BUILD_DIR=json_build
 
 PROCSRC="./$JSON_BUILD_DIR/*.o \
-	./src/haversine.cpp \
-	./src/point_types.cpp \
-	./src/processor/proc_argparser.cpp \
-	./src/processor/main.cpp "
+	./src/haversine.cc \
+	./src/point_types.cc \
+	./src/processor/proc_argparser.cc \
+	./src/processor/main.cc "
 PROCOUT=prochavr
 
 # MEMTESTER
@@ -113,7 +113,7 @@ if [[ $BASIC_PROFILING == true ]] || [[ $FULL_PROFILING == true ]]; then
 	cd ../
 
 	# REPETITION TESTING
-	REPTESTSRC="./src/repetition_testing/*.cpp ./$PROF_BUILD_DIR/*.o $ASM_LIB"
+	REPTESTSRC="./src/repetition_testing/*.cc ./$PROF_BUILD_DIR/*.o $ASM_LIB"
 	REPTESTOUT=reptest
 
 	(set -x ; $CXX $CFLAGS $REPTESTFLAGS $REPTESTSRC -o $REPTESTOUT)

haversine_02/src/repetition_testing/main.cc

@@ -5,6 +5,7 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#include <sys/mman.h>
 #include <unistd.h>
 
 #define ARR_LEN(ARR) sizeof(ARR) / sizeof(*ARR)
@@ -47,6 +48,8 @@ extern "C" void read_32x2_simd_offset(char *buffer, u64 size);
 extern "C" void read_32x2_simd_no_offset(char *buffer, u64 size);
 extern "C" void read_16x4_simd(char *buffer, u64 size);
 extern "C" void read_32x4_simd(char *buffer, u64 size);
+extern "C" void cache_test(char *buffer, u64 size, u64 mask);
+extern "C" void cache_test_unaligned(char *buffer, u64 size, u64 mask);
 
 void test_fread(reptester *tester, alloc_type type);
 void test_read(reptester *tester, alloc_type type);
@@ -89,6 +92,23 @@ void test_read_32x2_simd_offset(reptester *tester, alloc_type type);
 void test_read_32x2_simd_no_offset(reptester *tester, alloc_type type);
 void test_read_16x4_simd(reptester *tester, alloc_type type);
 void test_read_32x4_simd(reptester *tester, alloc_type type);
+void test_cache_test_16k(reptester *tester, alloc_type type);
+void test_cache_test_16k_unaligned(reptester *tester, alloc_type type);
+void test_cache_test_32k(reptester *tester, alloc_type type);
+void test_cache_test_32k_unaligned(reptester *tester, alloc_type type);
+void test_cache_test_64k(reptester *tester, alloc_type type);
+void test_cache_test_128k(reptester *tester, alloc_type type);
+void test_cache_test_256k(reptester *tester, alloc_type type);
+void test_cache_test_512k(reptester *tester, alloc_type type);
+void test_cache_test_1m(reptester *tester, alloc_type type);
+void test_cache_test_2m(reptester *tester, alloc_type type);
+void test_cache_test_4m(reptester *tester, alloc_type type);
+void test_cache_test_8m(reptester *tester, alloc_type type);
+void test_cache_test_16m(reptester *tester, alloc_type type);
+void test_cache_test_32m(reptester *tester, alloc_type type);
+void test_cache_test_64m(reptester *tester, alloc_type type);
+void test_cache_test_512m(reptester *tester, alloc_type type);
+void test_cache_test_full(reptester *tester, alloc_type type);
 u64 get_file_length(FILE *fp);
 
 int main(int argc, char *argv[]) {
@@ -118,20 +138,20 @@ int main(int argc, char *argv[]) {
       read_cpu_timer(),           // test_start_time
 
       1,                          // current_run
-			{
-				UINT64_MAX,               // min_time
-				0,                        // max_time
-				0,                        // avg_time
-				0,                        // total_time
-			},
-			{
-				0,                        // min_faults
-				0,                        // max_faults
-				0,                        // avg_faults
-				0,                        // total_bytes
-				0,                        // total_faults
-			},
-		  {},                         // results
+      {
+        UINT64_MAX,               // min_time
+        0,                        // max_time
+        0,                        // avg_time
+        0,                        // total_time
+      },
+      {
+        0,                        // min_faults
+        0,                        // max_faults
+        0,                        // avg_faults
+        0,                        // total_bytes
+        0,                        // total_faults
+      },
+      {},                         // results
   };
   // clang-format on
 
@@ -182,22 +202,58 @@ int main(int argc, char *argv[]) {
       // {{"READ 2x2", "READ 2x2 WITH MALLOC"}, test_read_2x2},
       // {{"READ 4x2", "READ 4x2 WITH MALLOC"}, test_read_4x2},
       // {{"READ 8x2", "READ 8x2 WITH MALLOC"}, test_read_8x2},
-      {{"READ 4x2_simd", "READ 4x2_simd WITH MALLOC"}, test_read_4x2_simd},
-      {{"READ 8x2_simd", "READ 8x2_simd WITH MALLOC"}, test_read_8x2_simd},
-      {{"READ 16x2_simd", "READ 16x2_simd WITH MALLOC"}, test_read_16x2_simd},
-      {{"READ 32x2_simd_offset", "READ 32x2_simd_offset WITH MALLOC"},
-       test_read_32x2_simd_offset},
-      {{"READ 32x2_simd_no_offset", "READ 32x2_simd_no_offset WITH MALLOC"},
-       test_read_32x2_simd_no_offset},
-      {{"READ 16x4_simd", "READ 16x4_simd WITH MALLOC"}, test_read_16x4_simd},
-      {{"READ 32x4_simd", "READ 32x4_simd WITH MALLOC"}, test_read_32x4_simd},
+      // {{"READ 4x2_simd", "READ 4x2_simd WITH MALLOC"}, test_read_4x2_simd},
+      // {{"READ 8x2_simd", "READ 8x2_simd WITH MALLOC"}, test_read_8x2_simd},
+      // {{"READ 16x2_simd", "READ 16x2_simd WITH MALLOC"},
+      // test_read_16x2_simd},
+      // {{"READ 32x2_simd_offset", "READ 32x2_simd_offset WITH MALLOC"},
+      //  test_read_32x2_simd_offset},
+      // {{"READ 32x2_simd_no_offset", "READ 32x2_simd_no_offset WITH MALLOC"},
+      //  test_read_32x2_simd_no_offset},
+      // {{"READ 16x4_simd", "READ 16x4_simd WITH MALLOC"},
+      // test_read_16x4_simd},
+      // {{"READ 32x4_simd", "READ 32x4_simd WITH MALLOC"},
+      // test_read_32x4_simd},
+      {{"CACHE TEST 16K", "CACHE TEST 16K WITH MALLOC"}, test_cache_test_16k},
+      {{"CACHE TEST 16K UNALIGNED", "CACHE TEST 16K UNALIGNED WITH MALLOC"},
+       test_cache_test_16k_unaligned},
+      // {{"CACHE TEST 32K", "CACHE TEST 32K WITH MALLOC"},
+      // test_cache_test_32k},
+      // {{"CACHE TEST 64K", "CACHE TEST 64K WITH MALLOC"},
+      // test_cache_test_64k},
+      // {{"CACHE TEST 128K", "CACHE TEST 128K WITH MALLOC"},
+      //  test_cache_test_128k},
+      // {{"CACHE TEST 256K", "CACHE TEST 256K WITH MALLOC"},
+      //  test_cache_test_256k},
+      // {{"CACHE TEST 512K", "CACHE TEST 512K WITH MALLOC"},
+      //  test_cache_test_512k},
+      // {{"CACHE TEST 1M", "CACHE TEST 1M WITH MALLOC"}, test_cache_test_1m},
+      // {{"CACHE TEST 2M", "CACHE TEST 2M WITH MALLOC"}, test_cache_test_2m},
+      // {{"CACHE TEST 4M", "CACHE TEST 4M WITH MALLOC"}, test_cache_test_4m},
+      // {{"CACHE TEST 8M", "CACHE TEST 8M WITH MALLOC"}, test_cache_test_8m},
+      // {{"CACHE TEST 16M", "CACHE TEST 16M WITH MALLOC"},
+      // test_cache_test_16m},
+      // {{"CACHE TEST 32M", "CACHE TEST 32M WITH MALLOC"},
+      // test_cache_test_32m},
+      // {{"CACHE TEST 64M", "CACHE TEST 64M WITH MALLOC"},
+      // test_cache_test_64m},
+      // {{"CACHE TEST 512M", "CACHE TEST 512M WITH MALLOC"},
+      //  test_cache_test_512m},
+      // {{"CACHE TEST FULL", "CACHE TEST FULL WITH MALLOC"},
+      //  test_cache_test_full},
   };
 
   tester.params.read_size = get_file_length(fp);
   tester.params.read_count = 1;
-  tester.params.buffer = (char *)malloc(tester.params.read_size + 1);
+  tester.params.buffer =
+      (char *)mmap(NULL, tester.params.read_size + 1, PROT_READ | PROT_WRITE,
+                   MAP_ANON | MAP_SHARED | MAP_NORESERVE, -1, 0);
   memset(tester.params.buffer, 0, tester.params.read_size + 1);
 
+  for (u64 i = 0; i < tester.params.read_size; ++i) {
+    tester.params.buffer[i] = (char)i;
+  }
+
   for (u64 i = 0; i < waves; ++i) {
     for (u64 j = 0; j < ARR_LEN(funcs); ++j) {
       for (u64 k = 0; k < ALLOC_TYPE_WITH_MALLOC; ++k) {
@@ -208,7 +264,7 @@ int main(int argc, char *argv[]) {
 
   fclose(fp);
 
-  free(tester.params.buffer);
+  munmap(tester.params.buffer, tester.params.read_size + 1);
 
   return 0;
 }
@@ -586,7 +642,6 @@ void test_align63_loop(reptester *tester, alloc_type type) {
 }
 
 void test_align75_loop(reptester *tester, alloc_type type) {
-
   u64 start = read_cpu_timer();
   u64 fault_count_start = page_fault_count();
 
@@ -612,7 +667,6 @@ void test_align75_loop(reptester *tester, alloc_type type) {
 }
 
 void test_align90_loop(reptester *tester, alloc_type type) {
-
   u64 start = read_cpu_timer();
   u64 fault_count_start = page_fault_count();
 
@@ -1112,7 +1166,6 @@ void test_read_4x2_simd(reptester *tester, alloc_type type) {
 }
 
 void test_read_8x2_simd(reptester *tester, alloc_type type) {
-
   u64 start = read_cpu_timer();
   u64 fault_count_start = page_fault_count();
 
@@ -1138,7 +1191,6 @@ void test_read_8x2_simd(reptester *tester, alloc_type type) {
 }
 
 void test_read_16x2_simd(reptester *tester, alloc_type type) {
-
   u64 start = read_cpu_timer();
   u64 fault_count_start = page_fault_count();
 
@@ -1264,6 +1316,431 @@ void test_read_32x4_simd(reptester *tester, alloc_type type) {
   handle_free(tester, type);
 }
 
+void test_cache_test_16k(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  cache_test(tester->params.buffer, total_size, 0x3fff);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_cache_test_16k_unaligned(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  cache_test_unaligned(tester->params.buffer, total_size, 0x3fff);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_cache_test_32k(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  cache_test(tester->params.buffer, total_size, 0x7fff);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_cache_test_32k_unaligned(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  cache_test_unaligned(tester->params.buffer, total_size, 0x7fff);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_cache_test_64k(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  cache_test(tester->params.buffer, total_size, 0xffff);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_cache_test_128k(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  cache_test(tester->params.buffer, total_size, 0x1ffff);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_cache_test_256k(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  cache_test(tester->params.buffer, total_size, 0x3ffff);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_cache_test_512k(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  cache_test(tester->params.buffer, total_size, 0x7ffff);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_cache_test_1m(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  cache_test(tester->params.buffer, total_size, 0xfffff);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_cache_test_2m(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  cache_test(tester->params.buffer, total_size, 0x1fffff);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_cache_test_4m(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  cache_test(tester->params.buffer, total_size, 0x3fffff);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_cache_test_8m(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  cache_test(tester->params.buffer, total_size, 0x7fffff);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_cache_test_16m(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  cache_test(tester->params.buffer, total_size, 0xffffff);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_cache_test_32m(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  cache_test(tester->params.buffer, total_size, 0x1ffffff);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_cache_test_64m(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  cache_test(tester->params.buffer, total_size, 0x3ffffff);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_cache_test_512m(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  cache_test(tester->params.buffer, total_size, 0x1fffffff);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
+void test_cache_test_full(reptester *tester, alloc_type type) {
+  u64 start = read_cpu_timer();
+  u64 fault_count_start = page_fault_count();
+
+  handle_alloc(tester, type);
+
+  u64 total_size = tester->params.read_size * tester->params.read_count;
+
+  cache_test(tester->params.buffer, total_size, 0xffffffffffffffff);
+
+  u64 fault_count_end = page_fault_count();
+  u64 end = read_cpu_timer();
+
+  u64 read_time = end - start;
+  u64 page_faults = fault_count_end - fault_count_start;
+
+  tester->results = {
+      total_size,
+      read_time,
+      page_faults,
+  };
+
+  handle_free(tester, type);
+}
+
 u64 get_file_length(FILE *fp) {
   if (!fp) {
     return 0;

haversine_02/src/repetition_testing/reptest_functions.asm

@@ -36,459 +36,494 @@ global read_32x2_simd_offset
 global read_32x2_simd_no_offset
 global read_16x4_simd
 global read_32x4_simd
+global cache_test ; Expects 3 inputs (pointer, read_count, mask)
+global cache_test_unaligned ; Expects 3 inputs (pointer, read_count, mask)
 
 mov_all_bytes_asm:
-	xor rax, rax
+  xor rax, rax
 
-	.loop:
-		mov BYTE [rdi + rax * 1], al
-		inc rax
-		cmp rsi, rax
-		jne .loop
+  .loop:
+    mov BYTE [rdi + rax * 1], al
+    inc rax
+    cmp rsi, rax
+    jne .loop
 
-	ret
+  ret
 
 nop_all_bytes_asm:
-	xor rax, rax
+  xor rax, rax
 
-	.loop:
-		db 0x0f, 0x1f, 0x00
-		inc rax
-		cmp rdi, rax
-		jne .loop
+  .loop:
+    db 0x0f, 0x1f, 0x00
+    inc rax
+    cmp rdi, rax
+    jne .loop
 
-	ret
+  ret
 
 nop_1x3_all_bytes_asm:
-	xor rax, rax
+  xor rax, rax
 
-	.loop:
-		nop
-		nop
-		nop
-		inc rax
-		cmp rdi, rax
-		jne .loop
+  .loop:
+    nop
+    nop
+    nop
+    inc rax
+    cmp rdi, rax
+    jne .loop
 
-	ret
+  ret
 
 nop_1x9_all_bytes_asm:
-	xor rax, rax
+  xor rax, rax
 
-	.loop:
-		nop
-		nop
-		nop
-		nop
-		nop
-		nop
-		nop
-		nop
-		nop
-		inc rax
-		cmp rdi, rax
-		jne .loop
+  .loop:
+    nop
+    nop
+    nop
+    nop
+    nop
+    nop
+    nop
+    nop
+    nop
+    inc rax
+    cmp rdi, rax
+    jne .loop
 
-	ret
+  ret
 
 inc_all_bytes_asm:
-	xor rax, rax
+  xor rax, rax
 
-	.loop:
-		inc rax
-		cmp rdi, rax
-		jne .loop
+  .loop:
+    inc rax
+    cmp rdi, rax
+    jne .loop
 
-	ret
+  ret
 
 dec_all_bytes_asm:
-	.loop:
-		dec rdi
-		jnz .loop
+  .loop:
+    dec rdi
+    jnz .loop
 
-	ret
+  ret
 
 align64_loop:
-	xor rax, rax
+  xor rax, rax
 
-	align 64
-	.loop:
-		inc rax
-		cmp rdi, rax
-		jne .loop
+  align 64
+  .loop:
+    inc rax
+    cmp rdi, rax
+    jne .loop
 
-	ret
+  ret
 
 align1_loop:
-	xor rax, rax
+  xor rax, rax
 
-	align 64
-	nop
-	.loop:
-		inc rax
-		cmp rdi, rax
-		jne .loop
+  align 64
+  nop
+  .loop:
+    inc rax
+    cmp rdi, rax
+    jne .loop
 
-	ret
+  ret
 
 align15_loop:
-	xor rax, rax
+  xor rax, rax
 
-	align 64
-	%rep 15
-	nop
-	%endrep
-	.loop:
-		inc rax
-		cmp rdi, rax
-		jne .loop
+  align 64
+  %rep 15
+  nop
+  %endrep
+  .loop:
+    inc rax
+    cmp rdi, rax
+    jne .loop
 
-	ret
+  ret
 
 align31_loop:
-	xor rax, rax
+  xor rax, rax
 
-	align 64
-	%rep 31
-	nop
-	%endrep
-	.loop:
-		inc rax
-		cmp rdi, rax
-		jne .loop
+  align 64
+  %rep 31
+  nop
+  %endrep
+  .loop:
+    inc rax
+    cmp rdi, rax
+    jne .loop
 
-	ret
+  ret
 
 align63_loop:
-	xor rax, rax
+  xor rax, rax
 
-	align 64
-	%rep 63
-	nop
-	%endrep
-	.loop:
-		inc rax
-		cmp rdi, rax
-		jne .loop
+  align 64
+  %rep 63
+  nop
+  %endrep
+  .loop:
+    inc rax
+    cmp rdi, rax
+    jne .loop
 
-	ret
+  ret
 
 align75_loop:
-	xor rax, rax
+  xor rax, rax
 
-	align 64
-	%rep 75
-	nop
-	%endrep
-	.loop:
-		inc rax
-		cmp rdi, rax
-		jne .loop
+  align 64
+  %rep 75
+  nop
+  %endrep
+  .loop:
+    inc rax
+    cmp rdi, rax
+    jne .loop
 
-	ret
+  ret
 
 align90_loop:
-	xor rax, rax
+  xor rax, rax
 
-	align 64
-	%rep 90
-	nop
-	%endrep
-	.loop:
-		inc rax
-		cmp rdi, rax
-		jne .loop
+  align 64
+  %rep 90
+  nop
+  %endrep
+  .loop:
+    inc rax
+    cmp rdi, rax
+    jne .loop
 
-	ret
+  ret
 
 align112_loop:
-	xor rax, rax
+  xor rax, rax
 
-	align 64
-	%rep 112
-	nop
-	%endrep
-	.loop:
-		inc rax
-		cmp rdi, rax
-		jne .loop
+  align 64
+  %rep 112
+  nop
+  %endrep
+  .loop:
+    inc rax
+    cmp rdi, rax
+    jne .loop
 
-	ret
+  ret
 
 rat_add:
-	mov rax, rdi
+  mov rax, rdi
 
-	.loop:
-		add rcx, 1
-		add rcx, 1
-		dec rax
-		jnz .loop
-	ret
+  .loop:
+    add rcx, 1
+    add rcx, 1
+    dec rax
+    jnz .loop
+  ret
 
 rat_mov_add:
-	mov rax, rdi
+  mov rax, rdi
 
-	.loop:
-		mov rcx, rax
-		add rcx, 1
-		mov rcx, rax
-		add rcx, 1
-		dec rax
-		jnz .loop
-	ret
+  .loop:
+    mov rcx, rax
+    add rcx, 1
+    mov rcx, rax
+    add rcx, 1
+    dec rax
+    jnz .loop
+  ret
 
 read_1:
-	align 64
-	.loop:
-		mov rax, [rdi]
-		sub rsi, 1
-		jnle .loop
+  align 64
+  .loop:
+    mov rax, [rdi]
+    sub rsi, 1
+    jnle .loop
 
-	ret
+  ret
 
 read_2:
-	align 64
-	.loop:
-		%rep 2
-		mov rax, [rdi]
-		%endrep
-		sub rsi, 2
-		jnle .loop
+  align 64
+  .loop:
+    %rep 2
+    mov rax, [rdi]
+    %endrep
+    sub rsi, 2
+    jnle .loop
 
-	ret
+  ret
 
 read_3:
-	align 64
-	.loop:
-		%rep 3
-		mov rax, [rdi]
-		%endrep
-		sub rsi, 3
-		jnle .loop
+  align 64
+  .loop:
+    %rep 3
+    mov rax, [rdi]
+    %endrep
+    sub rsi, 3
+    jnle .loop
 
-	ret
+  ret
 
 read_4:
-	align 64
-	.loop:
-		%rep 4
-		mov rax, [rdi]
-		%endrep
-		sub rsi, 4
-		jnle .loop
+  align 64
+  .loop:
+    %rep 4
+    mov rax, [rdi]
+    %endrep
+    sub rsi, 4
+    jnle .loop
 
-	ret
+  ret
 
 read_8:
-	align 64
-	.loop:
-		%rep 8
-		mov rax, [rdi]
-		%endrep
-		sub rsi, 8
-		jnle .loop
+  align 64
+  .loop:
+    %rep 8
+    mov rax, [rdi]
+    %endrep
+    sub rsi, 8
+    jnle .loop
 
-	ret
+  ret
 
 write_1:
-	align 64
-	.loop:
-		mov QWORD [rdi], 0
-		sub rsi, 1
-		jnle .loop
+  align 64
+  .loop:
+    mov QWORD [rdi], 0
+    sub rsi, 1
+    jnle .loop
 
-	ret
+  ret
 
 write_2:
-	align 64
-	.loop:
-		%rep 2
-		mov QWORD [rdi], 0
-		%endrep
-		sub rsi, 2
-		jnle .loop
+  align 64
+  .loop:
+    %rep 2
+    mov QWORD [rdi], 0
+    %endrep
+    sub rsi, 2
+    jnle .loop
 
-	ret
+  ret
 
 write_3:
-	align 64
-	.loop:
-		%rep 3
-		mov QWORD [rdi], 0
-		%endrep
-		sub rsi, 3
-		jnle .loop
+  align 64
+  .loop:
+    %rep 3
+    mov QWORD [rdi], 0
+    %endrep
+    sub rsi, 3
+    jnle .loop
 
-	ret
+  ret
 
 write_4:
-	align 64
-	.loop:
-		%rep 4
-		mov QWORD [rdi], 0
-		%endrep
-		sub rsi, 4
-		jnle .loop
+  align 64
+  .loop:
+    %rep 4
+    mov QWORD [rdi], 0
+    %endrep
+    sub rsi, 4
+    jnle .loop
 
-	ret
+  ret
 
 write_8:
-	align 64
-	.loop:
-		%rep 8
-		mov QWORD [rdi], 0
-		%endrep
-		sub rsi, 8
-		jnle .loop
+  align 64
+  .loop:
+    %rep 8
+    mov QWORD [rdi], 0
+    %endrep
+    sub rsi, 8
+    jnle .loop
 
-	ret
+  ret
 
 read_1x2_low:
-	align 64
-	.loop:
-		%rep 2
-		mov al, [rdi]
-		%endrep
-		sub rsi, 2
-		jnle .loop
+  align 64
+  .loop:
+    %rep 2
+    mov al, [rdi]
+    %endrep
+    sub rsi, 2
+    jnle .loop
 
-	ret
+  ret
 
 read_1x2_high:
-	align 64
-	.loop:
-		%rep 2
-		mov ah, [rdi]
-		%endrep
-		sub rsi, 2
-		jnle .loop
+  align 64
+  .loop:
+    %rep 2
+    mov ah, [rdi]
+    %endrep
+    sub rsi, 2
+    jnle .loop
 
-	ret
+  ret
 
 read_2x2:
-	align 64
-	.loop:
-		%rep 2
-		mov ax, [rdi]
-		%endrep
-		sub rsi, 2
-		jnle .loop
+  align 64
+  .loop:
+    %rep 2
+    mov ax, [rdi]
+    %endrep
+    sub rsi, 2
+    jnle .loop
 
-	ret
+  ret
 
 read_4x2:
-	align 64
-	.loop:
-		%rep 2
-		mov eax, [rdi]
-		%endrep
-		sub rsi, 2
-		jnle .loop
+  align 64
+  .loop:
+    %rep 2
+    mov eax, [rdi]
+    %endrep
+    sub rsi, 2
+    jnle .loop
 
-	ret
+  ret
 
 read_8x2:
-	align 64
-	.loop:
-		%rep 2
-		mov rax, [rdi]
-		%endrep
-		sub rsi, 2
-		jnle .loop
+  align 64
+  .loop:
+    %rep 2
+    mov rax, [rdi]
+    %endrep
+    sub rsi, 2
+    jnle .loop
 
-	ret
+  ret
 
 read_4x2_simd:
-	xor rax, rax
+  xor rax, rax
 
-	align 64
-	.loop:
-		mov r8d, [rdi]
-		mov r8d, [rdi + 4]
-		add rax, 8
-		cmp rax, rsi
-		jb .loop
+  align 64
+  .loop:
+    mov r8d, [rdi]
+    mov r8d, [rdi + 4]
+    add rax, 8
+    cmp rax, rsi
+    jb .loop
 
-	ret
+  ret
 
 read_8x2_simd:
-	xor rax, rax
+  xor rax, rax
 
-	align 64
-	.loop:
-		mov r8, [rdi]
-		mov r8, [rdi + 8]
-		add rax, 16
-		cmp rax, rsi
-		jb .loop
+  align 64
+  .loop:
+    mov r8, [rdi]
+    mov r8, [rdi + 8]
+    add rax, 16
+    cmp rax, rsi
+    jb .loop
 
-	ret
+  ret
 
 read_16x2_simd:
-	xor rax, rax
+  xor rax, rax
 
-	align 64
-	.loop:
-		vmovdqu xmm0, [rdi]
-		vmovdqu xmm0, [rdi + 16]
-		add rax, 32
-		cmp rax, rsi
-		jb .loop
+  align 64
+  .loop:
+    vmovdqu xmm0, [rdi]
+    vmovdqu xmm0, [rdi + 16]
+    add rax, 32
+    cmp rax, rsi
+    jb .loop
 
-	ret
+  ret
 
 read_32x2_simd_offset:
-	xor rax, rax
+  xor rax, rax
 
-	align 64
-	.loop:
-		vmovdqu ymm0, [rdi]
-		vmovdqu ymm0, [rdi + 32]
-		add rax, 64
-		cmp rax, rsi
-		jb .loop
+  align 64
+  .loop:
+    vmovdqu ymm0, [rdi]
+    vmovdqu ymm0, [rdi + 32]
+    add rax, 64
+    cmp rax, rsi
+    jb .loop
 
-	ret
+  ret
 
 read_32x2_simd_no_offset:
-	xor rax, rax
+  xor rax, rax
 
-	align 64
-	.loop:
-		vmovdqu ymm0, [rdi]
-		vmovdqu ymm0, [rdi]
-		add rax, 64
-		cmp rax, rsi
-		jb .loop
+  align 64
+  .loop:
+    vmovdqu ymm0, [rdi]
+    vmovdqu ymm0, [rdi]
+    add rax, 64
+    cmp rax, rsi
+    jb .loop
 
-	ret
+  ret
 
 read_16x4_simd:
-	xor rax, rax
+  xor rax, rax
 
-	align 64
-	.loop:
-		%rep 2
-		vmovdqu xmm0, [rdi]
-		vmovdqu xmm0, [rdi + 16]
-		%endrep
-		add rax, 64
-		cmp rax, rsi
-		jb .loop
+  align 64
+  .loop:
+    %rep 2
+    vmovdqu xmm0, [rdi]
+    vmovdqu xmm0, [rdi + 16]
+    %endrep
+    add rax, 64
+    cmp rax, rsi
+    jb .loop
 
-	ret
+  ret
 
 read_32x4_simd:
-	xor rax, rax
+  xor rax, rax
 
-	align 64
-	.loop:
-		%rep 2
-		vmovdqu ymm0, [rdi]
-		vmovdqu ymm0, [rdi]
-		%endrep
-		add rax, 128
-		cmp rax, rsi
-		jb .loop
+  align 64
+  .loop:
+    %rep 2
+    vmovdqu ymm0, [rdi]
+    vmovdqu ymm0, [rdi]
+    %endrep
+    add rax, 128
+    cmp rax, rsi
+    jb .loop
 
-	ret
+  ret
+
+cache_test:
+  xor r10, r10           ; Zero loop counter
+  mov rbx, rdi           ; Save original pointer
+  .loop:
+    add rdi, r10         ; Advance the pointer
+    add r10, 128         ; Increment loop counter
+    and r10, rdx         ; Mask offset
+    vmovdqu ymm0, [rdi + 0]
+    vmovdqu ymm1, [rdi + 32]
+    vmovdqu ymm2, [rdi + 64]
+    vmovdqu ymm3, [rdi + 96]
+    mov rdi, rbx         ; Restore original pointer
+    sub rsi, 128         ; Decrement count
+    ja .loop
+  ret
+
+cache_test_unaligned:
+  xor r10, r10           ; Zero loop counter
+	add rdi, 5             ; Unalign pointer
+  mov rbx, rdi           ; Save original pointer
+  .loop:
+    add rdi, r10         ; Advance the pointer
+    add r10, 128         ; Increment loop counter
+    and r10, rdx         ; Mask offset
+    vmovdqu ymm0, [rdi + 0]
+    vmovdqu ymm1, [rdi + 32]
+    vmovdqu ymm2, [rdi + 64]
+    vmovdqu ymm3, [rdi + 96]
+    mov rdi, rbx         ; Restore original pointer
+    sub rsi, 128         ; Decrement count
+    ja .loop
+  ret

haversine_02/src/repetition_testing/reptester.cc

@@ -3,6 +3,7 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#include <sys/mman.h>
 #include <sys/resource.h>
 #include <sys/time.h>
 
@@ -10,7 +11,9 @@ void handle_alloc(reptester *tester, alloc_type type) {
   switch (type) {
   case ALLOC_TYPE_WITH_MALLOC:
     if (!(tester->params.buffer)) {
-      tester->params.buffer = (char *)malloc(tester->params.read_size + 1);
+      tester->params.buffer = (char *)mmap(
+          NULL, tester->params.read_size + 1, PROT_READ | PROT_WRITE,
+          MAP_ANON | MAP_SHARED | MAP_NORESERVE, -1, 0);
       memset(tester->params.buffer, 0, tester->params.read_size + 1);
     }
 
@@ -24,7 +27,7 @@ void handle_free(reptester *tester, alloc_type type) {
   switch (type) {
   case ALLOC_TYPE_WITH_MALLOC:
     if (tester->params.buffer) {
-      free(tester->params.buffer);
+      munmap(tester->params.buffer, tester->params.read_size + 1);
       tester->params.buffer = NULL;
     }
 
@@ -58,7 +61,9 @@ void run_func_test(reptester *tester, reptest_func func, const char *func_name,
 
   if (type == ALLOC_TYPE_WITH_MALLOC) {
     buffer = tester->params.buffer;
-    tester->params.buffer = (char *)malloc(tester->params.read_size + 1);
+    tester->params.buffer =
+        (char *)mmap(NULL, tester->params.read_size + 1, PROT_READ | PROT_WRITE,
+                     MAP_ANON | MAP_SHARED | MAP_NORESERVE, -1, 0);
     memset(tester->params.buffer, 0, tester->params.read_size + 1);
   }
 
@@ -100,7 +105,7 @@ void run_func_test(reptester *tester, reptest_func func, const char *func_name,
   }
 
   if (type == ALLOC_TYPE_WITH_MALLOC) {
-    free(tester->params.buffer);
+    munmap(tester->params.buffer, tester->params.read_size + 1);
     tester->params.buffer = buffer;
   }