From a4324737d5e2c51c92add6a1f47c32f0548568e7 Mon Sep 17 00:00:00 2001
From: Stefan Arentz <stefan@arentz.ca>
Date: Thu, 21 Sep 2017 22:18:53 -0400
Subject: [PATCH] Performance improvements (#160)

---
 src/cpu.c       |  85 +++++++---------------------------------
 src/cpu.h       |   4 +-
 src/cpu_bench.c |  17 ++++++--
 src/cpu_test.c  |  25 +++++++-----
 src/mem.c       | 100 +++++++++++++++++++++++++++++++++++-------------
 src/mem.h       |   3 +-
 6 files changed, 121 insertions(+), 113 deletions(-)

diff --git a/src/cpu.c b/src/cpu.c
index 548bca2..c926b91 100644
--- a/src/cpu.c
+++ b/src/cpu.c
@@ -42,21 +42,21 @@
 // Stack management.
 
 void _cpu_push_byte(struct cpu_t *cpu, uint8_t b) {
-   cpu->page1[cpu->state.sp--] = b;
+   cpu->ram[0x0100 + cpu->state.sp--] = b;
 }
 
 void _cpu_push_word(struct cpu_t *cpu, uint16_t w) {
-   cpu->page1[cpu->state.sp--] = w >> 8;
-   cpu->page1[cpu->state.sp--] = w;
+   cpu->ram[0x0100 + cpu->state.sp--] = w >> 8;
+   cpu->ram[0x0100 + cpu->state.sp--] = w;
 }
 
 uint8_t _cpu_pull_byte(struct cpu_t *cpu) {
-   return cpu->page1[++cpu->state.sp];
+   return cpu->ram[0x0100 + ++cpu->state.sp];
 }
 
 uint16_t _cpu_pull_word(struct cpu_t *cpu) {
-   uint16_t w = (uint16_t) cpu->page1[++cpu->state.sp];
-   return w | ((uint16_t) cpu->page1[++cpu->state.sp] << 8);
+   uint16_t w = (uint16_t) cpu->ram[0x0100 + ++cpu->state.sp];
+   return w | ((uint16_t) cpu->ram[0x0100 + ++cpu->state.sp] << 8);
 }
 
 uint8_t _cpu_stack_free(struct cpu_t *cpu) {
@@ -95,43 +95,12 @@ void _cpu_set_status(struct cpu_t *cpu, uint8_t status) {
 }
 
 static int cpu_execute_instruction(struct cpu_t *cpu) {
-   /* Trace code - Refactor into its own function or module */
-   char trace_instruction[256];
-   char trace_state[256];
-   char trace_stack[256];
-
-   if (cpu->trace) {
-      cpu_format_instruction(cpu, trace_instruction);
-   }
-
-   /* Fetch instruction */
+   // Fetch instruction
    struct cpu_instruction_t *i = &cpu->instructions[mem_get_byte(cpu, cpu->state.pc)];
-   if (i->name[0] == '?') {
-      if (cpu->strict) {
-         return EWM_CPU_ERR_UNIMPLEMENTED_INSTRUCTION;
-      }
-   }
 
-   // If strict mode and if we need the stack, check if that works out
-   if (cpu->strict && i->stack != 0) {
-      if (i->stack > 0) {
-         if (_cpu_stack_free(cpu) < i->stack) {
-            return EWM_CPU_ERR_STACK_OVERFLOW;
-         }
-      } else {
-         if (_cpu_stack_used(cpu) < -(i->stack)) {
-            return EWM_CPU_ERR_STACK_UNDERFLOW;
-         }
-      }
-   }
-
-   /* Remember the PC since some instructions modify it */
+   // Remember and advance the pc
    uint16_t pc = cpu->state.pc;
-
-   /* Advance PC */
-   if (pc == cpu->state.pc) {
-      cpu->state.pc += i->bytes;
-   }
+   cpu->state.pc += i->bytes;
 
    if (i->lua_before_handler != LUA_NOREF) {
       lua_rawgeti(cpu->lua->state, LUA_REGISTRYINDEX, i->lua_before_handler);
@@ -186,27 +155,6 @@ static int cpu_execute_instruction(struct cpu_t *cpu) {
       }
    }
 
-   if (cpu->trace) {
-      cpu_format_state(cpu, trace_state);
-      cpu_format_stack(cpu, trace_stack);
-
-      char bytes[10];
-      switch (i->bytes) {
-         case 1:
-            snprintf(bytes, sizeof bytes, "%.2X", mem_get_byte(cpu, pc));
-            break;
-         case 2:
-            snprintf(bytes, sizeof bytes, "%.2X %.2X", mem_get_byte(cpu, pc), mem_get_byte(cpu, pc+1));
-            break;
-         case 3:
-            snprintf(bytes, sizeof bytes, "%.2X %.2X %.2X", mem_get_byte(cpu, pc), mem_get_byte(cpu, pc+1), mem_get_byte(cpu, pc+2));
-            break;
-      }
-
-      fprintf(cpu->trace, "%.4X: %-8s  %-14s  %-20s  %s\n",
-              pc, bytes, trace_instruction, trace_state, trace_stack);
-   }
-
    cpu->counter += i->cycles;
 
    return i->cycles;
@@ -408,19 +356,14 @@ struct mem_t *cpu_add_iom(struct cpu_t *cpu, uint16_t start, uint16_t end, void
 // memory regions.
 
 void cpu_optimize_memory(struct cpu_t *cpu) {
-   struct mem_t *page0 = cpu_mem_for_page(cpu, 0);
-   if (page0 == NULL || (page0->flags != (MEM_FLAGS_READ | MEM_FLAGS_WRITE))) {
-      printf("[CPU] Cannot find rw memory region that handles Page 0\n");
+   struct mem_t *zp= cpu_mem_for_page(cpu, 0);
+   if (zp == NULL || (zp->flags != (MEM_FLAGS_READ | MEM_FLAGS_WRITE)) || zp->end < 0x1ff) {
+      printf("[CPU] Cannot find rw memory region that covers 0x0000 to 0x01ff\n");
       exit(1);
    }
-   cpu->page0 = page0->obj + (0x0000 - page0->start);
 
-   struct mem_t *page1 = cpu_mem_for_page(cpu, 1);
-   if (page1 == NULL || (page1->flags != (MEM_FLAGS_READ | MEM_FLAGS_WRITE))) {
-      printf("[CPU] Cannot find rw memory region that handles Page 1\n");
-      exit(1);
-   }
-   cpu->page1 = page1->obj + (0x0100 - page1->start);
+   cpu->ram = zp->obj;
+   cpu->ram_size = zp->end;
 }
 
 void cpu_strict(struct cpu_t *cpu, bool strict) {
diff --git a/src/cpu.h b/src/cpu.h
index 21a1124..e252300 100644
--- a/src/cpu.h
+++ b/src/cpu.h
@@ -56,8 +56,8 @@ struct cpu_t {
    struct cpu_instruction_t *instructions;
    uint64_t counter;
 
-   uint8_t *page0;
-   uint8_t *page1;
+   uint8_t *ram;
+   size_t ram_size;
 
    struct ewm_lua_t *lua;
 };
diff --git a/src/cpu_bench.c b/src/cpu_bench.c
index 216512f..08db3fc 100644
--- a/src/cpu_bench.c
+++ b/src/cpu_bench.c
@@ -22,6 +22,7 @@
 
 #include <stdlib.h>
 #include <stdint.h>
+#include <string.h>
 #include <time.h>
 
 #include "cpu.h"
@@ -29,7 +30,7 @@
 #include "mem.h"
 #include "utl.h"
 
-#define CPU_BENCH_ITERATIONS (100 * 1000 * 1000)
+#define CPU_BENCH_ITERATIONS (10 * 1000 * 1000)
 
 void test(struct cpu_t *cpu, uint8_t opcode) {
    uint64_t runs[3];
@@ -83,7 +84,17 @@ int main(int argc, char **argv) {
    cpu_add_ram_data(cpu, 0, 0xffff, malloc(0xffff));
    cpu_reset(cpu);
 
-   for (int opcode = 0; opcode <= 255; opcode++) {
-      test(cpu, opcode);
+   if (argc > 1) {
+      for (int i = 1; i < argc; i++) {
+         for (int opcode = 0; opcode <= 255; opcode++) {
+            if (strcmp(cpu->instructions[opcode].name, argv[i]) == 0) {
+               test(cpu, opcode);
+            }
+         }
+      }
+   } else {
+      for (int opcode = 0; opcode <= 255; opcode++) {
+         test(cpu, opcode);
+      }
    }
 }
diff --git a/src/cpu_test.c b/src/cpu_test.c
index b86b168..3611306 100644
--- a/src/cpu_test.c
+++ b/src/cpu_test.c
@@ -32,21 +32,21 @@
 #include "utl.h"
 #include "lua.h"
 
-int test(int model, uint16_t start_addr, uint16_t success_addr, char *rom_path) {
+int test(int model, uint16_t start_addr, uint16_t success_addr, char *rom_path, int with_lua) {
    struct cpu_t *cpu = cpu_create(model);
    cpu_add_ram_file(cpu, 0x0000, rom_path);
    cpu_reset(cpu);
    cpu->state.pc = start_addr;
 
-#if 1
-   cpu->lua = ewm_lua_create();
-   ewm_cpu_init_lua(cpu, cpu->lua);
+   if (with_lua) {
+      cpu->lua = ewm_lua_create();
+      ewm_cpu_init_lua(cpu, cpu->lua);
 
-   if (ewm_lua_load_script(cpu->lua, "cpu_test.lua") != 0) {
-      printf("Lua script failed to load\n"); // TODO Move errors reporting into C code
-      exit(1);
+      if (ewm_lua_load_script(cpu->lua, "cpu_test.lua") != 0) {
+         printf("Lua script failed to load\n"); // TODO Move errors reporting into C code
+         exit(1);
+      }
    }
-#endif
 
    uint16_t last_pc = cpu->state.pc;
 
@@ -114,7 +114,12 @@ int test(int model, uint16_t start_addr, uint16_t success_addr, char *rom_path)
 
 int main(int argc, char **argv) {
    fprintf(stderr, "TEST Running 6502 tests\n");
-   test(EWM_CPU_MODEL_6502,  0x0400, 0x3399, "rom/6502_functional_test.bin");
+   test(EWM_CPU_MODEL_6502,  0x0400, 0x3399, "rom/6502_functional_test.bin", 0);
    fprintf(stderr, "TEST Running 65C02 tests\n");
-   test(EWM_CPU_MODEL_65C02, 0x0400, 0x24a8, "rom/65C02_extended_opcodes_test.bin");
+   test(EWM_CPU_MODEL_65C02, 0x0400, 0x24a8, "rom/65C02_extended_opcodes_test.bin", 0);
+
+   fprintf(stderr, "TEST Running 6502 tests - With Lua\n");
+   test(EWM_CPU_MODEL_6502,  0x0400, 0x3399, "rom/6502_functional_test.bin", 1);
+   fprintf(stderr, "TEST Running 65C02 tests - With Lua\n");
+   test(EWM_CPU_MODEL_65C02, 0x0400, 0x24a8, "rom/65C02_extended_opcodes_test.bin", 1);
 }
diff --git a/src/mem.c b/src/mem.c
index 15b9be0..618a621 100644
--- a/src/mem.c
+++ b/src/mem.c
@@ -33,6 +33,10 @@
 // take more time but do the right thing.
 
 uint8_t mem_get_byte(struct cpu_t *cpu, uint16_t addr) {
+   if (addr < cpu->ram_size) {
+      return cpu->ram[addr];
+   }
+
    struct mem_t *mem = cpu->mem;
    while (mem != NULL) {
       if (mem->enabled && addr >= mem->start && addr <= mem->end) {
@@ -42,15 +46,15 @@ uint8_t mem_get_byte(struct cpu_t *cpu, uint16_t addr) {
       }
       mem = mem->next;
    }
-
-   if (cpu->strict) {
-      // TODO: Signal an error about reading non-existent memory?
-   }
-
    return 0;
 }
 
 void mem_set_byte(struct cpu_t *cpu, uint16_t addr, uint8_t v) {
+   if (addr < cpu->ram_size) {
+      cpu->ram[addr] = v;
+      return;
+   }
+
    struct mem_t *mem = cpu->mem;
    while (mem != NULL) {
       if (mem->enabled && addr >= mem->start && addr <= mem->end) {
@@ -61,90 +65,134 @@ void mem_set_byte(struct cpu_t *cpu, uint16_t addr, uint8_t v) {
       }
       mem = mem->next;
    }
-
-   if (cpu->strict) {
-      // TODO: Signal an error about writing non-existent memory?
-   }
 }
 
 // Getters
 
 uint8_t mem_get_byte_abs(struct cpu_t *cpu, uint16_t addr) {
-  return mem_get_byte(cpu, addr);
+   if (addr < cpu->ram_size) {
+      return cpu->ram[addr];
+   }
+   return mem_get_byte(cpu, addr);
 }
 
 uint8_t mem_get_byte_absx(struct cpu_t *cpu, uint16_t addr) {
+   if (addr < cpu->ram_size) {
+      return cpu->ram[addr + cpu->state.x];
+   }
    return mem_get_byte(cpu, addr + cpu->state.x);
 }
 
 uint8_t mem_get_byte_absy(struct cpu_t *cpu, uint16_t addr) {
-  return mem_get_byte(cpu, addr + cpu->state.y);
+   return cpu->ram[addr + cpu->state.y];
 }
 
 uint8_t mem_get_byte_zpg(struct cpu_t *cpu, uint8_t addr) {
-   return cpu->page0[addr];
+   return cpu->ram[addr];
 }
 
 uint8_t mem_get_byte_zpgx(struct cpu_t *cpu, uint8_t addr) {
-   return cpu->page0[((uint16_t) addr + cpu->state.x) & 0x00ff];
+   return cpu->ram[((uint16_t) addr + cpu->state.x) & 0x00ff];
 }
 
 uint8_t mem_get_byte_zpgy(struct cpu_t *cpu, uint8_t addr) {
-   return cpu->page0[((uint16_t) addr + cpu->state.y) & 0x00ff];
+   return cpu->ram[((uint16_t) addr + cpu->state.y) & 0x00ff];
 }
 
 uint8_t mem_get_byte_indx(struct cpu_t *cpu, uint8_t addr) {
-   return mem_get_byte(cpu, (((uint16_t) cpu->page0[((uint16_t)addr+1+cpu->state.x)&0x00ff] << 8) | (uint16_t) cpu->page0[((uint16_t) addr+cpu->state.x) & 0x00ff]));
+   uint16_t a = *((uint16_t*) &cpu->ram[(addr + cpu->state.x) & 0x0ff]);
+   if (a < cpu->ram_size) {
+      return cpu->ram[a];
+   }
+   return mem_get_byte(cpu, a);
 }
 
 uint8_t mem_get_byte_indy(struct cpu_t *cpu, uint8_t addr) {
-   return mem_get_byte(cpu, (((uint16_t) cpu->page0[addr+1] << 8) | (uint16_t) cpu->page0[addr]) + cpu->state.y);
+   uint16_t a = *((uint16_t*) &cpu->ram[addr]) + cpu->state.y;
+   if (a < cpu->ram_size) {
+      return cpu->ram[a];
+   }
+   return mem_get_byte(cpu, a);
 }
 
 uint8_t mem_get_byte_ind(struct cpu_t *cpu, uint8_t addr) {
-   return mem_get_byte(cpu, ((uint16_t) cpu->page0[addr+1] << 8) | (uint16_t) cpu->page0[addr]);
+   uint16_t a = *((uint16_t*) &cpu->ram[addr]);
+   if (a < cpu->ram_size) {
+      return cpu->ram[a];
+   }
+   return mem_get_byte(cpu, a);
 }
 
 uint16_t mem_get_word(struct cpu_t *cpu, uint16_t addr) {
+  if (addr < cpu->ram_size) {
+     return *((uint16_t*) &cpu->ram[addr]);
+  }
   return ((uint16_t) mem_get_byte(cpu, addr+1) << 8) | (uint16_t) mem_get_byte(cpu, addr);
 }
 
 // Setters
 
 void mem_set_byte_zpg(struct cpu_t *cpu, uint8_t addr, uint8_t v) {
-   cpu->page0[addr] = v;
+   cpu->ram[addr] = v;
 }
 
 void mem_set_byte_zpgx(struct cpu_t *cpu, uint8_t addr, uint8_t v) {
-   cpu->page0[((uint16_t) addr + cpu->state.x) & 0x00ff] = v;
+   cpu->ram[((uint16_t) addr + cpu->state.x) & 0x00ff] = v;
 }
 
 void mem_set_byte_zpgy(struct cpu_t *cpu, uint8_t addr, uint8_t v) {
-   cpu->page0[((uint16_t) addr + cpu->state.y) & 0x00ff] = v;
+   cpu->ram[((uint16_t) addr + cpu->state.y) & 0x00ff] = v;
 }
 
 void mem_set_byte_abs(struct cpu_t *cpu, uint16_t addr, uint8_t v) {
-  mem_set_byte(cpu, addr, v);
+   if (addr < cpu->ram_size) {
+      cpu->ram[addr] = v;
+      return;
+   }
+   mem_set_byte(cpu, addr, v);
 }
 
 void mem_set_byte_absx(struct cpu_t *cpu, uint16_t addr, uint8_t v) {
-  mem_set_byte(cpu, addr+cpu->state.x, v);
+   if (addr < cpu->ram_size) {
+      cpu->ram[addr + cpu->state.x] = v;
+      return;
+   }
+   mem_set_byte(cpu, addr + cpu->state.x, v);
 }
 
 void mem_set_byte_absy(struct cpu_t *cpu, uint16_t addr, uint8_t v) {
-  mem_set_byte(cpu, addr+cpu->state.y, v);
+   if (addr < cpu->ram_size) {
+      cpu->ram[addr + cpu->state.y] = v;
+      return;
+   }
+   mem_set_byte(cpu, addr + cpu->state.y, v);
 }
 
 void mem_set_byte_indx(struct cpu_t *cpu, uint8_t addr, uint8_t v) {
-   mem_set_byte(cpu, (((uint16_t) cpu->page0[((uint16_t)addr+1+cpu->state.x)&0x00ff] << 8) | (uint16_t) cpu->page0[((uint16_t) addr+cpu->state.x) & 0x00ff]), v);
+   uint16_t a = *((uint16_t*) &cpu->ram[(addr + cpu->state.x) & 0x0ff]);
+   if (a < cpu->ram_size) {
+      cpu->ram[a] = v;
+      return;
+   }
+   mem_set_byte(cpu, a, v);
 }
 
 void mem_set_byte_indy(struct cpu_t *cpu, uint8_t addr, uint8_t v) {
-   mem_set_byte(cpu, (((uint16_t) cpu->page0[addr+1] << 8) | (uint16_t) cpu->page0[addr]) + cpu->state.y, v);
+   uint16_t a = *((uint16_t*) &cpu->ram[addr]) + cpu->state.y;
+   if (a < cpu->ram_size) {
+      cpu->ram[a] = v;
+      return;
+   }
+   mem_set_byte(cpu, a, v);
 }
 
 void mem_set_byte_ind(struct cpu_t *cpu, uint8_t addr, uint8_t v) {
-   mem_set_byte(cpu, (((uint16_t) cpu->page0[addr+1] << 8) | (uint16_t) cpu->page0[addr]), v);
+   uint16_t a = *((uint16_t*) &cpu->ram[addr]);
+   if (a < cpu->ram_size) {
+      cpu->ram[a] = v;
+      return;
+   }
+   mem_set_byte(cpu, a, v);
 }
 
 void mem_set_word(struct cpu_t *cpu, uint16_t addr, uint16_t v) {
diff --git a/src/mem.h b/src/mem.h
index b3b7417..4470834 100644
--- a/src/mem.h
+++ b/src/mem.h
@@ -30,6 +30,8 @@ struct cpu_t;
 typedef uint8_t (*mem_mod_t)(struct cpu_t *cpu, uint8_t b);
 
 uint8_t mem_get_byte(struct cpu_t *cpu, uint16_t addr);
+void mem_set_byte(struct cpu_t *cpu, uint16_t addr, uint8_t v);
+
 uint8_t mem_get_byte_abs(struct cpu_t *cpu, uint16_t addr);
 uint8_t mem_get_byte_absx(struct cpu_t *cpu, uint16_t addr);
 uint8_t mem_get_byte_absy(struct cpu_t *cpu, uint16_t addr);
@@ -40,7 +42,6 @@ uint8_t mem_get_byte_indx(struct cpu_t *cpu, uint8_t addr);
 uint8_t mem_get_byte_indy(struct cpu_t *cpu, uint8_t addr);
 uint8_t mem_get_byte_ind(struct cpu_t *cpu, uint8_t addr);
 
-void mem_set_byte(struct cpu_t *cpu, uint16_t addr, uint8_t v);
 void mem_set_byte_zpg(struct cpu_t *cpu, uint8_t addr, uint8_t v);
 void mem_set_byte_zpgx(struct cpu_t *cpu, uint8_t addr, uint8_t v);
 void mem_set_byte_zpgy(struct cpu_t *cpu, uint8_t addr, uint8_t v);