adapt emulator for generator changes:

- new instructions (push, pop, call, ret)
- new pseudo-registers (bp, sp, rv)
- new binary format
- dynamic register file size
- removed obsolete instructions

14 files changed, 450 insertions, 219 deletions

diff --git a/src/emu/asm.c b/src/emu/asm.c
index caca72a..8e39202 100644
--- a/src/emu/asm.c
+++ b/src/emu/asm.c
@@ -5,28 +5,49 @@
 #include "log.h"
 #include "opc.h"
 
+#define REGa(x)   (((x) & 0x1F) << 21)
+#define REGb(x)   (((x) & 0x1F) << 16)
+#define REGc(x)   (((x) & 0x1F) << 11)
+#define IMMc(x)   ((x) & 0xFFFF)
+
 uint32_t compile(const char *line)
 {
-	char mnem[4];
+	char mnem[5];
 	int32_t a = 0, b = 0, c = 0;
+	static size_t regs = 0;
+
+	/* register file size */
+	if (sscanf(line, ".REGS %d", &a) == 1) {
+		regs = a;
+		return a;
+	}
+
+	/* ADD, SUB, MUL, DIV, MOD, AND, OR */
+	if (sscanf(line, "%s r%d, r%d, r%d", mnem, &a, &b, &c) == 4)
+		return mnemonic2opc(mnem) | REGa(a) | REGb(b) | REGc(c);
+
+	/* LW, SW */
+	if (sscanf(line, "%s r%d, r%d, %d", mnem, &a, &b, &c) == 4)
+		return mnemonic2opc(mnem) | REGa(a) | REGb(b) | IMMc(c);
 
-	/* arithmetic & logic */
-	if (sscanf(line, "%3s r%2d, r%2d, r%2d", mnem, &a, &b, &c) == 4)
-		return mnemonic2opc(mnem) | ((a & 0x1F) << 21) | ((b & 0x1F) << 16) | ((c & 0x1F) << 11);
+	/* CMP */
+	if (sscanf(line, "%s r%d, r%d", mnem, &a, &b) == 3)
+		return mnemonic2opc(mnem) | REGa(a) | REGb(b);
 
-	/* load/store & branch */
-	if (sscanf(line, "%3s r%2d, r%2d, %d", mnem, &a, &b, &c) == 4)
-		return mnemonic2opc(mnem) | ((a & 0x1F) << 21) | ((b & 0x1F) << 16) | (c & 0xFFFF);
+	/* MOV, BEZ */
+	if (sscanf(line, "%s r%2d, %d", mnem, &a, &c) == 3)
+		return mnemonic2opc(mnem) | REGa(a) | IMMc(c);
 
-	if (sscanf(line, "%3s r%2d, %d", mnem, &b, &c) == 3)
-		return mnemonic2opc(mnem) | ((b & 0x1F) << 16) | (c & 0xFFFF);
+	/* EQ, NE, LT, LE, GE, GT, PUSH, POP */
+	if (sscanf(line, "%s r%d", mnem, &a) == 2)
+		return mnemonic2opc(mnem) | REGa(a);
 
-	/* jump */
-	if (sscanf(line, "%3s r%2d", mnem, &a) == 2)
-		return mnemonic2opc(mnem) | ((a & 0x1F) << 21);
+	/* JMP, CALL */
+	if (sscanf(line, "%s %d", mnem, &c) == 2)
+		return mnemonic2opc(mnem) | IMMc(c);
 
-	/* misc */
-	if (sscanf(line, "%3s", mnem) == 1)
+	/* RET, SYS */
+	if (sscanf(line, "%s", mnem) == 1)
 		return mnemonic2opc(mnem);
 
 	return 0xFFFFFFFF;
diff --git a/src/emu/cpu.c b/src/emu/cpu.c
index 7e62ffb..f32bb7f 100644
--- a/src/emu/cpu.c
+++ b/src/emu/cpu.c
@@ -11,16 +11,25 @@
 #include "opc.h"
 #include "syscall.h"
 
-// program counter
-uint32_t PC;
+/* stack pointer register*/
+uint32_t SP;
 
-// status bits
+/* base pointer register */
+uint32_t BP;
+
+/* return value register */
+uint32_t RV;
+
+/* instruction pointer */
+uint32_t IP;
+
+/* status bits */
 bool N, Z;
 
-// 32 general purpose registers
-uint32_t GPR[32];
+/* general purpose registers */
+uint32_t *GPR;
 
-/* extract operands from the instruction regiter */
+/* extract operands from the instruction register */
 static inline
 uint32_t __OPEXT(uint32_t IR, uint8_t start, uint8_t length)
 {
@@ -41,16 +50,19 @@ void trap(int num)
 {
 	switch (num) {
 	case TRP_UNALIGNED:
-		raise(SIGSEGV);
+		fprintf(stderr, "trap: unaligned memory access\n");
+		abort();
 		break;
 	case TRP_DIVBYZERO:
-		raise(SIGFPE);
+		fprintf(stderr, "trap: division by zero\n");
+		abort();
 		break;
 	case TRP_SYSCALL:
 		do_syscall();
 		break;
 	case TRP_ILL:
-		raise(SIGILL);
+		fprintf(stderr, "trap: illegal instruction\n");
+		abort();
 		break;
 	}
 }
@@ -62,17 +74,17 @@ void execute(uint32_t IR)
 	int32_t a, b, c;
 
 	if (opcode < OPC_MOV) {
-		/* arithmetic & logic */
+		/* arithmetic, logic, comparison */
 		a = REGa;
 		b = REGb;
 		c = REGc;
 
-		debug("PC@%#08x: %-3s r%i, r%i, r%i", PC,
+		debug("IP@%#08x: %-4s r%i, r%i, r%i", IP,
 				opc2mnemonic(IR), a, b, c);
 	}
 
-	else if (opcode < OPC_J) {
-		/* load/store & branch */
+	else if (opcode < OPC_BEZ) {
+		/* load/store */
 		a = REGa;
 		b = REGb;
 		c = IMMc;
@@ -81,40 +93,59 @@ void execute(uint32_t IR)
 		if (c >= 0x8000)
 			c -= 0x10000;
 
-		if (opcode < OPC_LB) {
-			debug("PC@%#08x: %-3s r%i, %i", PC,
-					opc2mnemonic(IR), b, c);
-		} else {
-			debug("PC@%#08x: %-3s r%i, r%i, %i", PC,
+		if (opcode < OPC_LW) {
+			debug("IP@%#08x: %-4s r%i, %i", IP,
+					opc2mnemonic(IR), a, c);
+		} else if (opcode < OPC_PUSH) {
+			debug("IP@%#08x: %-4s r%i, r%i, %i", IP,
 					opc2mnemonic(IR), a, b, c);
+		} else {
+			debug("IP@%#08x: %-4s r%i", IP,
+					opc2mnemonic(IR), a);
 		}
 	}
 
 	else if (opcode < OPC_SYS) {
 		/* jump */
 		a = REGa;
-		b = c = 0;
+		b = 0;
+		c = IMMc;
 
-		debug("PC@%#08x: %-3s r%i", PC,
-				opc2mnemonic(IR), a);
+		/* sign extension */
+		if (c >= 0x8000)
+			c -= 0x10000;
+
+		switch (opcode) {
+		case OPC_BEZ:
+			debug("IP@%#08x: %-4s r%i, %i", IP,
+					opc2mnemonic(IR), a, c);
+			break;
+		case OPC_JMP:
+			debug("IP@%#08x: %-4s %i", IP,
+					opc2mnemonic(IR), c);
+			break;
+		case OPC_CALL:
+			debug("IP@%#08x: %-4s %i", IP,
+					opc2mnemonic(IR), c);
+			break;
+		case OPC_RET:
+			debug("IP@%#08x: %-4s", IP,
+					opc2mnemonic(IR));
+			break;
+		}
 	}
 
 	else {
 		/* misc */
 		a = b = c = 0;
 
-		debug("PC@%#08x: %-3s", PC,
+		debug("IP@%#08x: %-3s", IP,
 				opc2mnemonic(IR));
 	}
 
 	// make sure r0 is zero
 	GPR[0] = 0;
 
-	// buffer for load/store instructions
-	uint8_t tmp8;
-	uint16_t tmp16;
-	uint32_t tmp32;
-
 	switch (opcode) {
 	case OPC_ADD:
 		// XXX: signed/unsigned?
@@ -139,69 +170,16 @@ void execute(uint32_t IR)
 		// XXX: signed/unsigned?
 		GPR[a] = GPR[b] % GPR[c];
 		break;
-	case OPC_SHL:
-		// XXX: signed/unsigned?
-		GPR[a] = GPR[b] << GPR[c];
-		break;
-	case OPC_SHR:
-		// XXX: signed/unsigned?
-		GPR[a] = GPR[b] >> GPR[c];
-		break;
 	case OPC_AND:
 		GPR[a] = GPR[b] & GPR[c];
 		break;
 	case OPC_OR:
 		GPR[a] = GPR[b] | GPR[c];
 		break;
-	case OPC_XOR:
-		GPR[a] = GPR[b] ^ GPR[c];
-		break;
-	case OPC_NOR:
-		GPR[a] = ~(GPR[b] & GPR[c]);
-		break;
-	case OPC_MOV:
-		GPR[b] = c;
-		break;
-	case OPC_LB:
-		memcpy(&tmp8, &MEM[GPR[a] + c], sizeof(uint8_t));
-		GPR[b] = tmp8;
-		break;
-	case OPC_LH:
-		if ((GPR[a] + c) & 0x1)
-			trap(TRP_UNALIGNED);
-		memcpy(&tmp16, &MEM[GPR[a] + c], sizeof(uint16_t));
-		GPR[b] = tmp16;
-		break;
-	case OPC_LW:
-		if ((GPR[a] + c) & 0x2)
-			trap(TRP_UNALIGNED);
-		memcpy(&tmp32, &MEM[GPR[a] + c], sizeof(uint32_t));
-		GPR[b] = tmp32;
-		break;
-	case OPC_SB:
-		tmp8 = GPR[b];
-		memcpy(&MEM[GPR[a] + c], &tmp8, sizeof(uint8_t));
-		break;
-	case OPC_SH:
-		if ((GPR[a] + c) & 0x1)
-			trap(TRP_UNALIGNED);
-		tmp16 = GPR[b];
-		memcpy(&MEM[GPR[a] + c], &tmp16, sizeof(uint16_t));
-		break;
-	case OPC_SW:
-		if ((GPR[a] + c) & 0x2)
-			trap(TRP_UNALIGNED);
-		tmp32 = GPR[b];
-		memcpy(&MEM[GPR[a] + c], &tmp32, sizeof(uint32_t));
-		break;
 	case OPC_CMP:
 		Z = (GPR[a] == GPR[b]);
 		N = (GPR[a] < GPR[b]);
 		break;
-	case OPC_BEZ:
-		if (GPR[a] == 0)
-			PC += c * sizeof(uint32_t);
-		break;
 	case OPC_EQ:
 		GPR[a] = Z;
 		break;
@@ -219,12 +197,37 @@ void execute(uint32_t IR)
 		break;
 	case OPC_GT:
 		GPR[a] = !Z && !N;
-	case OPC_J:
-		PC = GPR[a];
+	case OPC_MOV:
+		GPR[a] = c;
 		break;
-	case OPC_JAL:
-		GPR[31] = PC + sizeof(uint32_t);
-		PC = GPR[a];
+	case OPC_LW:
+		GPR[b] = load(GPR[a] + c);
+		break;
+	case OPC_SW:
+		store(GPR[a] + c, GPR[b]);
+		break;
+	case OPC_PUSH:
+		push(GPR[a]);
+		break;
+	case OPC_POP:
+		GPR[a] = pop();
+		break;
+	case OPC_BEZ:
+		if (GPR[a] == 0) {
+			IP += c * 4;
+			return;
+		}
+		break;
+	case OPC_JMP:
+		IP += c * 4;
+		return;
+	case OPC_CALL:
+		push(IP + 4);
+		IP += c * 4;
+		return;
+	case OPC_RET:
+		IP = pop();
+		return;
 	case OPC_SYS:
 		trap(TRP_SYSCALL);
 		break;
@@ -232,4 +235,6 @@ void execute(uint32_t IR)
 		/* illegal instruction */
 		trap(TRP_ILL);
 	}
+
+	IP += 4;
 }
diff --git a/src/emu/cpu.h b/src/emu/cpu.h
index 43ccf00..4867a8a 100644
--- a/src/emu/cpu.h
+++ b/src/emu/cpu.h
@@ -14,14 +14,23 @@ enum {
 
 void trap(int num);
 
-/* program counter */
-extern uint32_t PC;
+/* stack pointer */
+extern uint32_t SP;
+
+/* base pointer */
+extern uint32_t BP;
+
+/* return value */
+extern uint32_t RV;
+
+/* instruction pointer */
+extern uint32_t IP;
 
 /* status bits */
 extern bool N, Z;
 
-/* 32 general purpose registers */
-extern uint32_t GPR[32];
+/* general purpose registers */
+extern uint32_t *GPR;
 
 /* main cpu execution function */
 void execute(uint32_t IR);
diff --git a/src/emu/mem.c b/src/emu/mem.c
index e102eaf..cdbf2ae 100644
--- a/src/emu/mem.c
+++ b/src/emu/mem.c
@@ -1,3 +1,37 @@
+#include <stdint.h>
+#include <sys/types.h>
+#include <string.h>
+
+#include "cpu.h"
 #include "mem.h"
 
-uint8_t MEM[4096];
+uint8_t *MEM;
+
+void store(size_t addr, uint32_t value)
+{
+	if (addr & 0x2)
+		trap(TRP_UNALIGNED);
+	memcpy(&MEM[addr], &value, 4);
+}
+
+uint32_t load(size_t addr)
+{
+	uint32_t tmp32;
+	if (addr & 0x2)
+		trap(TRP_UNALIGNED);
+	memcpy(&tmp32, &MEM[addr], 4);
+	return tmp32;
+}
+
+void push(uint32_t value)
+{
+	GPR[SP] -= 4;
+	store(GPR[SP], value);
+}
+
+uint32_t pop(void)
+{
+	uint32_t tmp32 = load(GPR[SP]);
+	GPR[SP] += 4;
+	return tmp32;
+}
diff --git a/src/emu/mem.h b/src/emu/mem.h
index 34b13fa..985a838 100644
--- a/src/emu/mem.h
+++ b/src/emu/mem.h
@@ -2,7 +2,14 @@
 #define _MEM_H
 
 #include <stdint.h>
+#include <sys/types.h>
 
-extern uint8_t MEM[];
+extern uint8_t *MEM;
+
+void store(size_t addr, uint32_t value);
+uint32_t load(size_t addr);
+
+void push(uint32_t value);
+uint32_t pop(void);
 
 #endif
diff --git a/src/emu/opc.c b/src/emu/opc.c
index fa46b26..3d62479 100644
--- a/src/emu/opc.c
+++ b/src/emu/opc.c
@@ -9,35 +9,30 @@ typedef struct opc_mapping {
 } opc_mapping_t;
 
 opc_mapping_t opc_map[] = {
-	{ "ADD", OPC_ADD },
-	{ "SUB", OPC_SUB },
-	{ "MUL", OPC_MUL },
-	{ "DIV", OPC_DIV },
-	{ "MOD", OPC_MOD },
-	{ "SHL", OPC_SHL },
-	{ "SHR", OPC_SHR },
-	{ "AND", OPC_AND },
-	{ "OR",  OPC_OR },
-	{ "XOR", OPC_XOR },
-	{ "NOR", OPC_NOR },
-	{ "MOV", OPC_MOV },
-	{ "LB",  OPC_LB },
-	{ "LH",  OPC_LH },
-	{ "LW",  OPC_LW },
-	{ "SB",  OPC_SB },
-	{ "SH",  OPC_SH },
-	{ "SW",  OPC_SW },
-	{ "CMP", OPC_CMP },
-	{ "BEZ", OPC_BEZ },
-	{ "EQ",  OPC_EQ },
-	{ "NE",  OPC_NE },
-	{ "LT",  OPC_LT },
-	{ "GE",  OPC_GE },
-	{ "LE",  OPC_LE },
-	{ "GT",  OPC_GT },
-	{ "J",   OPC_J },
-	{ "JAL", OPC_JAL },
-	{ "SYS", OPC_SYS },
+	{ "ADD",  OPC_ADD },
+	{ "SUB",  OPC_SUB },
+	{ "MUL",  OPC_MUL },
+	{ "DIV",  OPC_DIV },
+	{ "MOD",  OPC_MOD },
+	{ "AND",  OPC_AND },
+	{ "OR",   OPC_OR },
+	{ "CMP",  OPC_CMP },
+	{ "EQ",   OPC_EQ },
+	{ "NE",   OPC_NE },
+	{ "LT",   OPC_LT },
+	{ "GE",   OPC_GE },
+	{ "LE",   OPC_LE },
+	{ "GT",   OPC_GT },
+	{ "MOV",  OPC_MOV },
+	{ "LW",   OPC_LW },
+	{ "SW",   OPC_SW },
+	{ "PUSH", OPC_PUSH },
+	{ "POP",  OPC_POP },
+	{ "BEZ",  OPC_BEZ },
+	{ "JMP",  OPC_JMP },
+	{ "CALL", OPC_CALL },
+	{ "RET",  OPC_RET },
+	{ "SYS",  OPC_SYS },
 	{ NULL,  0 }
 };
 
diff --git a/src/emu/opc.h b/src/emu/opc.h
index b766afa..179f7b2 100644
--- a/src/emu/opc.h
+++ b/src/emu/opc.h
@@ -10,12 +10,12 @@
  *     |000|xxx|aaaaa|bbbbb|ccccc|00000000000|
  *  logic:
  *     |001|xxx|aaaaa|bbbbb|ccccc|00000000000|
+ *  comparison:
+ *     |010|xxx|aaaaa|bbbbb|0000000000000000|
  *  load & store:
- *     |010|xxx|aaaaa|bbbbb|cccccccccccccccc|
- *  branch:
  *     |011|xxx|aaaaa|bbbbb|cccccccccccccccc|
  *  jump:
- *     |100|xxx|aaaaa|000000000000000000000|
+ *     |100|xxx|aaaaa|00000|cccccccccccccccc|
  *  misc:
  *     |111|xxx|??????????????????????????|
  *
@@ -27,37 +27,32 @@
 #define OPC_MUL  002
 #define OPC_DIV  003
 #define OPC_MOD  004
-#define OPC_SHL  005
-#define OPC_SHR  006
 
 /* logic */
 #define OPC_AND  010
 #define OPC_OR   011
-#define OPC_XOR  012
-#define OPC_NOR  013
 
-/* load & store */
-#define OPC_MOV  020
-#define OPC_LB   021
-#define OPC_LH   022
-#define OPC_LW   023
-#define OPC_SB   024
-#define OPC_SH   025
-#define OPC_SW   026
+/* comparison */
+#define OPC_CMP  020
+#define OPC_EQ   021
+#define OPC_NE   022
+#define OPC_LT   023
+#define OPC_LE   024
+#define OPC_GE   025
+#define OPC_GT   026
 
-/* branch instructions */
-#define OPC_CMP  030
-#define OPC_BEZ  031
-#define OPC_EQ   032
-#define OPC_NE   033
-#define OPC_LT   034
-#define OPC_LE   035
-#define OPC_GE   036
-#define OPC_GT   037
+/* load & store */
+#define OPC_MOV  030
+#define OPC_LW   031
+#define OPC_SW   032
+#define OPC_PUSH 033
+#define OPC_POP  034
 
 /* jump instructions */
-#define OPC_J    040
-#define OPC_JAL  041
+#define OPC_BEZ  040
+#define OPC_JMP  041
+#define OPC_CALL 042
+#define OPC_RET  043
 
 /* misc */
 #define OPC_SYS  070
diff --git a/src/emu/riscas.c b/src/emu/riscas.c
index e4180b2..7a4af47 100644
--- a/src/emu/riscas.c
+++ b/src/emu/riscas.c
@@ -33,7 +33,7 @@ int main(int argc, char *argv[])
 		if (IR == 0xFFFFFFFF)
 			die("illegal instruction: %s", line);
 
-		fwrite(&IR, sizeof(uint32_t), 1, pfd);
+		fwrite(&IR, 4, 1, pfd);
 	}
 
 	return 0;
diff --git a/src/emu/risci.c b/src/emu/risci.c
index 8c4e30f..704ece8 100644
--- a/src/emu/risci.c
+++ b/src/emu/risci.c
@@ -1,9 +1,11 @@
 #include <unistd.h>
 #include <stdlib.h>
 #include <stdint.h>
+#include <inttypes.h>
 #include <stdbool.h>
 #include <stdio.h>
 #include <signal.h>
+#include <assert.h>
 #include <fcntl.h>
 #include <sys/stat.h>
 #include <readline/readline.h>
@@ -11,13 +13,12 @@
 #include "asm.h"
 #include "cpu.h"
 #include "log.h"
+#include "mem.h"
 #include "opc.h"
 
 bool is_debug = false;
-bool is_interactive = false;
-
-/* global program buffer */
-uint8_t *P = NULL;
+size_t mem_size = 64 * 1024 * 1024;
+size_t reg_size = 0;
 
 static
 void usage(int rc)
@@ -27,27 +28,11 @@ void usage(int rc)
 }
 
 static
-void signal_handler(int sig)
+void dump_registers(void)
 {
-	switch (sig) {
-	case SIGILL:
-		/* SIGILL is raised by the cpu for an unknown instruction */
-		error("ERROR: illegal instruction.");
-		if (!is_interactive)
-			exit(-1);
-		break;
-	case SIGSEGV:
-		/* SIGSEGV is raised for unaligned memory access */
-		error("ERROR: unaligned memory access.");
-		if (!is_interactive)
-			exit(-1);
-		break;
-	case SIGFPE:
-		/* SIGFPE is raised by devision with zero */
-		error("ERROR: division by zero.");
-		exit(-1);
-		break;
-	}
+	fprintf(stderr, "REGISTER DUMP:\n");
+	for (size_t i = 0; i < reg_size; i++)
+		fprintf(stderr, "\tr%zu = %"PRIu32"\n", i, GPR[i]);
 }
 
 static
@@ -57,31 +42,34 @@ void read_program(const char *program)
 	if (lstat(program, &sb) == -1)
 		pdie("cannot stat program");
 
-	if (sb.st_size % sizeof(uint32_t))
-		die("program does not align to op-code size of %u bytes", sizeof(uint32_t));
+	if (sb.st_size % 4)
+		die("program does not align to op-code size of 4 bytes");
 
 	int pfd;
 	if ((pfd = open(program, O_RDONLY)) == -1)
 		pdie("could not open program");
 
-	P = malloc(sb.st_size + sizeof(uint32_t));
-	if (read(pfd, P, sb.st_size) != sb.st_size)
-		die("premature end of program");
+	/* initialize register file */
+	assert(read(pfd, &reg_size, 4) == 4);
+	reg_size += 4;
+	GPR = calloc(reg_size, 4);
 
-	memset(P + sb.st_size, 0xFF, sizeof(uint32_t));
+	/* initialize pseudo registers */
+	BP = reg_size - 3;
+	SP = reg_size - 2;
+	RV = reg_size - 1;
+	GPR[BP] = GPR[SP] = mem_size;
+	GPR[RV] = 0;
+
+	assert(read(pfd, MEM, sb.st_size - 4) == sb.st_size - 4);
+	memset(MEM + sb.st_size + 4, 0xFF, 4);
 }
 
 static
 uint32_t next_instruction(void)
 {
-	if (is_interactive) {
-		/* read next instruction from stdin */
-		printf("%03d", PC/4);
-		return compile(readline("> "));
-	}
-
 	uint32_t tmp;
-	memcpy(&tmp, &P[PC], sizeof(uint32_t));
+	memcpy(&tmp, &MEM[IP], 4);
 	return tmp;
 }
 
@@ -89,15 +77,15 @@ int main(int argc, char *argv[])
 {
 	int ch;
 
-	while ((ch = getopt(argc, argv, "dhi")) != -1) {
+	while ((ch = getopt(argc, argv, "dhm")) != -1) {
 		switch (ch) {
 		case 'd':
 			is_debug = true;
 			break;
 		case 'h':
 			usage(EXIT_SUCCESS);
-		case 'i':
-			is_interactive = true;
+		case 'm':
+			mem_size = atoi(optarg) * 4;
 			break;
 		case '?':
 		default:
@@ -108,26 +96,21 @@ int main(int argc, char *argv[])
 	argc -= optind;
 	argv += optind;
 
-	/* catch cpu signal traps */
-	signal(SIGILL,  signal_handler);
-	signal(SIGFPE,  signal_handler);
-	signal(SIGSEGV, signal_handler);
-
-	/* load program from file if we're not in interactive mode */
-	if (!is_interactive) {
-		if (argc < 1)
-			usage(EXIT_FAILURE);
+	/* reserve memory */
+	MEM = calloc(mem_size, sizeof(char));
 
-		read_program(argv[0]);
-	}
+	/* load program from file */
+	if (argc < 1)
+		usage(EXIT_FAILURE);
+	read_program(argv[0]);
 
 	/* reset program counter to first instruction */
-	PC = 0;
+	IP = 0;
 
 	/* start instruction loop */
 	while (1) {
+		//dump_registers();
 		execute(next_instruction());
-		PC += 4;
 	}
 
 	/* not reached, program is terminated by signal traps from the cpu */
diff --git a/src/emu/syscall.c b/src/emu/syscall.c
index fe0d83e..625bbf4 100644
--- a/src/emu/syscall.c
+++ b/src/emu/syscall.c
@@ -5,19 +5,21 @@
 #include "mem.h"
 #include "syscall.h"
 
+#define arg(n) load(GPR[SP] + 4*n)
+
 void do_syscall(void)
 {
-	switch (GPR[1]) {
+	switch (arg(0)) {
 	case SYS_exit:
-		exit(GPR[2]);
+		exit(arg(1));
 		break;
 	case SYS_read:
-		GPR[2] = read(GPR[2], &MEM[GPR[3]], GPR[4]);
+		push(read(arg(1), &MEM[arg(2)], arg(3)));
 		break;
 	case SYS_write:
-		GPR[2] = write(GPR[2], &MEM[GPR[3]], GPR[4]);
+		push(write(arg(1), &MEM[arg(2)], arg(3)));
 		break;
 	default:
-		GPR[2] = -1;
+		push(-1);
 	}
 }
diff --git a/src/emu/test1 b/src/emu/test1
new file mode 100644
index 0000000..e2cbe0a
--- /dev/null
+++ b/src/emu/test1
diff --git a/src/emu/test1.S b/src/emu/test1.S
new file mode 100644
index 0000000..1feae7a
--- /dev/null
+++ b/src/emu/test1.S
@@ -0,0 +1,126 @@
+.REGS 23
+JMP 53
+PUSH r24
+ADD r24, r0, r25
+PUSH r1
+PUSH r2
+PUSH r3
+PUSH r4
+PUSH r5
+PUSH r6
+PUSH r7
+PUSH r8
+PUSH r9
+PUSH r10
+LW r24, r2, 8
+MOV r3, 2
+CMP r2, r3
+LT  r1
+BEZ r1, 4
+MOV r4, 1
+ADD r26, r0, r4
+JMP 20
+LW r24, r5, 8
+MOV r6, 1
+SUB r5, r5, r6
+PUSH r5
+CALL -24
+ADD r7, r0, r26
+POP r0
+LW r24, r8, 8
+MOV r9, 2
+SUB r8, r8, r9
+PUSH r8
+CALL -31
+ADD r10, r0, r26
+POP r0
+ADD r7, r7, r10
+ADD r26, r0, r7
+JMP 3
+ADD r26, r0, r0
+JMP 1
+POP r10
+POP r9
+POP r8
+POP r7
+POP r6
+POP r5
+POP r4
+POP r3
+POP r2
+POP r1
+ADD r25, r0, r24
+POP r24
+RET
+PUSH r24
+ADD r24, r0, r25
+PUSH r0
+PUSH r0
+PUSH r0
+PUSH r11
+PUSH r12
+PUSH r13
+PUSH r14
+PUSH r15
+PUSH r16
+PUSH r17
+PUSH r18
+PUSH r19
+PUSH r20
+PUSH r21
+PUSH r22
+PUSH r23
+MOV r11, 0
+SW r24, r11, -4
+MOV r12, 0
+SW r24, r12, -8
+LW r24, r14, -8
+MOV r15, 10
+CMP r14, r15
+LT  r13
+BEZ r13, 14
+LW r24, r16, -4
+LW r24, r17, -8
+PUSH r17
+CALL -82
+ADD r18, r0, r26
+POP r0
+ADD r16, r16, r18
+SW r24, r16, -4
+LW r24, r19, -8
+MOV r20, 1
+ADD r19, r19, r20
+SW r24, r19, -8
+JMP -17
+MOV r21, 3
+PUSH r21
+CALL -94
+ADD r22, r0, r26
+POP r0
+SW r24, r22, -12
+LW r24, r23, -4
+ADD r26, r0, r23
+JMP 3
+ADD r26, r0, r0
+JMP 1
+POP r0
+POP r0
+POP r0
+POP r23
+POP r22
+POP r21
+POP r20
+POP r19
+POP r18
+POP r17
+POP r16
+POP r15
+POP r14
+POP r13
+POP r12
+POP r11
+ADD r25, r0, r24
+POP r24
+PUSH r26
+PUSH r0
+SYS
diff --git a/src/emu/test4 b/src/emu/test4
new file mode 100644
index 0000000..6219e58
--- /dev/null
+++ b/src/emu/test4
diff --git a/src/emu/test4.S b/src/emu/test4.S
new file mode 100644
index 0000000..08e5375
--- /dev/null
+++ b/src/emu/test4.S
@@ -0,0 +1,54 @@
+.REGS 7
+JMP 17
+PUSH r8
+ADD r8, r0, r9
+PUSH r1
+PUSH r2
+LW r8, r1, 8
+MOV r2, 2
+MUL r1, r1, r2
+ADD r10, r0, r1
+JMP 3
+ADD r10, r0, r0
+JMP 1
+POP r2
+POP r1
+ADD r9, r0, r8
+POP r8
+RET
+PUSH r8
+ADD r8, r0, r9
+PUSH r0
+PUSH r0
+PUSH r3
+PUSH r4
+PUSH r5
+PUSH r6
+PUSH r7
+MOV r3, 10
+MOV r4, 5
+ADD r3, r3, r4
+SW r8, r3, -4
+LW r8, r5, -4
+PUSH r5
+CALL -31
+ADD r6, r0, r10
+POP r0
+SW r8, r6, -8
+LW r8, r7, -8
+ADD r10, r0, r7
+JMP 3
+ADD r10, r0, r0
+JMP 1
+POP r0
+POP r0
+POP r7
+POP r6
+POP r5
+POP r4
+POP r3
+ADD r9, r0, r8
+POP r8
+PUSH r10
+PUSH r0
+SYS