add recursion test, add noops for branch delay slots

include/mips.h

@@ -304,7 +304,8 @@ MIPS_INS(JAL, .op = MIPS_OP_JAL)
 
 /* JALR - jump and link register */
 #define MIPS_FUNCT_JALR    0b001001
-MIPS_INS(JALR, .op = MIPS_OP_SPECIAL, .funct = MIPS_FUNCT_JALR)
+MIPS_INS(JALR, .rd = MIPS_REG_RA, .op = MIPS_OP_SPECIAL,
+	 .funct = MIPS_FUNCT_JALR)
 
 /* JALX - jump and link exchange */
 #define MIPS_OP_JALX       0b011101

lib/runtime.asm

@@ -23,6 +23,7 @@ _start:
 
 	# call main
 	jal main
+	nop
 
 	# exit
 	move	$a0, $v0

masm/parse.c

@@ -889,6 +889,17 @@ static int parse_pseudo_move(struct parser *parser, struct ins_expr *expr)
 	return M_SUCCESS;
 }
 
+static int parse_pseudo_nop(struct parser *parser, struct ins_expr *expr)
+{
+	(void) parser;
+
+	expr->ins_len = 1;
+	expr->ins[0] = mips_instructions[MIPS_INS_SLL];
+	expr->ref[0].type = R_MIPS_NONE;
+
+	return M_SUCCESS;
+}
+
 static int parse_pseudo_instruction(struct parser *parser,
 				    struct ins_expr *expr,
 				    struct token ident)
@@ -911,6 +922,10 @@ static int parse_pseudo_instruction(struct parser *parser,
 		res = parse_pseudo_la(parser, expr);
 	else CHK(move)
 		res = parse_pseudo_move(parser, expr);
+	else CHK(nop)
+		res = parse_pseudo_nop(parser, expr);
+
+	#undef CHK
 
 	if (res) {
 		// reset on fail

mld/link.c

@@ -12,8 +12,6 @@
 #include "link.h"
 #include "mips.h"
 
-#define SEC_ALIGN 0x1000
-
 static int load_objects(struct linker *linker)
 {
 	int max_entries = linker->args->in_count + 1;
@@ -168,7 +166,7 @@ static int relocate_symbol(struct linker *linker, struct object *obj,
 			   struct symbol_table *symtab, const Elf32_Sym *sym)
 {
 	size_t shndx = B16(sym->st_shndx);
-	if (shndx == 0)
+	if (shndx == 0 || shndx == SHN_ABS)
 		return M_SUCCESS; // ignore this symbol
 
 	// find the given section
@@ -184,11 +182,9 @@ static int relocate_symbol(struct linker *linker, struct object *obj,
 		return M_ERROR;
 	}
 
-	Elf32_Shdr const *shdr = &obj->shdr[shndx];
 	struct segment *sec = NULL;
 	for (size_t i = 0; i < obj->phdr_len; i++) {
-		Elf32_Phdr *temp = &obj->phdr[i];
-		if (PHDR_SHDR_MATCH(temp, shdr)) {
+		if (obj->phdr_to_shdr_mapping[i] == shndx) {
 			sec = &obj->segments[i];
 			break;
 		}
@@ -219,6 +215,20 @@ static int relocate_symbol(struct linker *linker, struct object *obj,
 	new.st_shndx = B16(new_shndx);
 	new.st_size = 0;
 
+	// rename symbol if its name is empty to perm placeholder
+	//if (B32(sym->st_name) == 0) {
+	//	#define __MAX 512
+	//	char name[__MAX];
+	//	memset(name, 0, __MAX);
+	//	strcat(name, "__sym");
+	//	snprintf(name + strlen(name), __MAX - strlen(name), "%d",
+	//		B32(sym->st_value));
+	//	if (strtab_push(linker->symtab.strtab, name, &str_off))
+	//		return M_ERROR;
+	//	new.st_name = B32(str_off);
+	//	#undef __MAX
+	//}
+
 	if (symtab_get(&linker->symtab, NULL, name, obj->index) == M_SUCCESS) {
 		ERROR("cannot link doubly defiend symbol '%s'", name);
 		return M_ERROR;
@@ -447,33 +457,38 @@ static int relocate_instruction_rela(struct linker *linker,
 		// 16bit absolute
 		if (vaddr_abs > UINT16_MAX)
 			warn = true;
-		ins.immd = (uint16_t)vaddr_abs;
+		ins.immd += (uint16_t)vaddr_abs;
 		break;
 	case R_MIPS_PC16:
 		// 16bit relative shifted
 		if (vaddr_rel > INT16_MAX || vaddr_rel < INT16_MIN)
 			warn = true;
-		ins.offset = vaddr_rel;
+		ins.offset += vaddr_rel;
 		break;
 	case R_MIPS_26:
+	case R_MIPS_JALR:
 		// 26bit absolute shifted
 		if (vaddr_abs >= (1 << 25))
 			warn = true;
-		ins.target = (vaddr_abs & 0x0FFFFFFF) >> 2;
+		ins.target += (vaddr_abs & 0x0FFFFFFF) >> 2;
 		break;
 	case R_MIPS_PC26_S2:
 		// 26bit relative shifted
 		if (vaddr_rel >= (1 << 24) || -vaddr_rel > (1 << 24))
 			warn = true;
-		ins.offs26 = vaddr_rel;
+		ins.offs26 += vaddr_rel;
 		break;
 	case R_MIPS_LO16:
 		// lo 16bit absolute
-		ins.immd = (uint16_t)(vaddr_abs & 0xFFFF);
+		ins.immd += (uint16_t)(vaddr_abs & 0xFFFF);
 		break;
 	case R_MIPS_HI16:
 		// hi 16bit absolute
-		ins.immd = (uint16_t)(vaddr_abs >> 16);
+		ins.immd += (uint16_t)(vaddr_abs >> 16);
+		break;
+	case R_MIPS_32:
+		// 32bit absolute
+		ins.raw = vaddr_abs;
 		break;
 	default:
 		ERROR("do not know how do handle relocation type [%d]", typ);
@@ -596,9 +611,6 @@ static void update_offsets(struct linker *linker)
 
 static int write_file(struct linker *linker)
 {
-	extern char *current_file;
-	current_file = linker->args->out_file;
-
 	int fd = open(linker->args->out_file, O_RDWR | O_CREAT, 0711);
 	if (fd < 0) {
 		PERROR("cannot write");
@@ -739,12 +751,16 @@ static void linker_free(struct linker *linker)
 
 int link_files(struct linker_arguments args) {
 	struct linker linker;
+	extern char *current_file;
+
 
 	int res = M_SUCCESS;
 	if (res == M_SUCCESS)
 		res = linker_init(&linker, &args);
 	if (res == M_SUCCESS)
 		res = load_objects(&linker);
+
+	current_file = args.out_file;
 	if (res == M_SUCCESS)
 		res = relocate_segments(&linker);
 	if (res == M_SUCCESS)

mld/link.h

@@ -35,6 +35,9 @@
 #define TEXT_VADDR_MIN 0x00400000
 #define DATA_VADDR_MIN 0x10000000
 
+// alignment of a section
+#define SEC_ALIGN 0x1000
+
 // pre define
 struct linker;
 struct object;
@@ -42,6 +45,7 @@ struct segment;
 struct string_table;
 struct symbol_table;
 struct symbol_table_mapping;
+struct glboff_table;
 
 ///
 /// relocation table
@@ -220,6 +224,10 @@ struct object {
 	// program table
 	Elf32_Phdr *phdr;
 	size_t phdr_len;
+	bool phdr_local; // if phdr was created though malloc
+
+	// phdr <=> shdr mappings
+	uint32_t *phdr_to_shdr_mapping;
 
 	// object meta
 	const char *name;

mld/obj.c

@@ -49,10 +49,10 @@ static int load_ehdr(struct object *object)
 	EHDR_ASSERT(type,      sizeof(Elf32_Half));
 	EHDR_ASSERT(machine,   sizeof(Elf32_Half));
 	EHDR_ASSERT(version,   sizeof(Elf32_Word));
-	EHDR_ASSERT(flags,     sizeof(Elf32_Word));
+	// EHDR_ASSERT(flags,     sizeof(Elf32_Word));
 	EHDR_ASSERT(ehsize,    sizeof(Elf32_Half));
-	EHDR_ASSERT(phentsize, sizeof(Elf32_Half));
-	EHDR_ASSERT(shentsize, sizeof(Elf32_Half));
+	// EHDR_ASSERT(phentsize, sizeof(Elf32_Half));
+	// EHDR_ASSERT(shentsize, sizeof(Elf32_Half));
 
 	#undef EHDR_ASSERT
 
@@ -81,23 +81,94 @@ static int load_shdr(struct object *object)
 	return M_SUCCESS;
 }
 
+/**
+ * Create the phdr
+ */
+static int create_phdr(struct object *object)
+{
+	uint32_t entries = 0;
+	for (uint32_t i = 0; i < object->shdr_len; i++) {
+		Elf32_Shdr *hdr = &object->shdr[i];
+		uint32_t type = B32(hdr->sh_type);
+
+		if (type != SHT_PROGBITS && type != SHT_NOBITS)
+			continue;
+
+		entries += 1;
+	}
+
+	Elf32_Phdr *phdr = malloc(entries * sizeof(Elf32_Phdr));
+	if (phdr == NULL) {
+		PERROR("cannot alloc");
+		return M_ERROR;
+	}
+
+	object->phdr = phdr;
+	object->phdr_len = entries;
+	object->phdr_local = true;
+
+	uint32_t *mapping = malloc(entries * sizeof(uint32_t));
+	if (mapping == NULL) {
+		PERROR("cannot alloc");
+		return M_ERROR;
+	}
+
+	object->phdr_to_shdr_mapping = mapping;
+
+	uint32_t index = 0;
+	for (uint32_t i = 0; i < object->shdr_len; i++) {
+		Elf32_Shdr *hdr = &object->shdr[i];
+		uint32_t type = B32(hdr->sh_type);
+
+		if (type != SHT_PROGBITS && type != SHT_NOBITS)
+			continue;
+		mapping[index] = i;
+		phdr[index++] = (Elf32_Phdr) {
+			.p_type = B32(PT_LOAD),
+			.p_flags = B32(
+				// execute
+				((B32(hdr->sh_flags) & SHF_EXECINSTR)
+					? PF_X : 0) |
+				// write
+				((B32(hdr->sh_flags) & SHF_WRITE)
+					? PF_W : 0) |
+				// read
+				((B32(hdr->sh_flags) & SHF_ALLOC)
+					? PF_R : 0)
+			),
+			.p_offset = hdr->sh_offset,
+			.p_vaddr = hdr->sh_addr,
+			.p_paddr = hdr->sh_addr,
+			.p_filesz = hdr->sh_size,
+			.p_memsz = hdr->sh_size,
+			.p_align = B32(SEC_ALIGN),
+		};
+	}
+
+	return M_SUCCESS;
+}
+
 /**
  * Map the phdr
  */
 static int load_phdr(struct object *object)
 {
-	size_t phdr_len = B16(object->ehdr->e_phentsize) *
-		B16(object->ehdr->e_phnum);
-	size_t phdr_off = B32(object->ehdr->e_phoff);
-	object->phdr = (Elf32_Phdr *) (object->mapped + phdr_off);
-	object->phdr_len = B16(object->ehdr->e_phnum);
+	//size_t phdr_len = B16(object->ehdr->e_phentsize) *
+	//	B16(object->ehdr->e_phnum);
 
-	if (BOUND_CHK(object, phdr_len, phdr_off)) {
-		ERROR("cannot read phdr");
-		return M_ERROR;
-	}
+	//if (phdr_len < 1)
+	return create_phdr(object);
 
-	return M_SUCCESS;
+	//size_t phdr_off = B32(object->ehdr->e_phoff);
+	//object->phdr = (Elf32_Phdr *) (object->mapped + phdr_off);
+	//object->phdr_len = B16(object->ehdr->e_phnum);
+
+	//if (BOUND_CHK(object, phdr_len, phdr_off)) {
+	//	ERROR("cannot read phdr");
+	//	return M_ERROR;
+	//}
+
+	//return M_SUCCESS;
 }
 
 /**
@@ -231,7 +302,7 @@ static int load_shstrtab(struct object *object)
  */
 static int load_segments(struct object *object)
 {
-	object->segment_len = B16(object->ehdr->e_phnum);
+	object->segment_len = object->phdr_len;
 	object->segments = malloc(sizeof(struct segment) *
 				object->segment_len);
 
@@ -287,6 +358,8 @@ int object_load(struct object *object, char *path, uint32_t index)
 	object->mapped = NULL;
 	object->name = path;
 	object->index = index;
+	object->phdr_local = false;
+	object->phdr_to_shdr_mapping = NULL;
 
 	/** load the file */
 	if (map_file(object, path))
@@ -331,6 +404,10 @@ void object_free(struct object *obj)
 		free(obj->strtabs);
 	if (obj->segments != NULL)
 		free(obj->segments);
+	if (obj->phdr_local)
+		free(obj->phdr);
+	if (obj->phdr_to_shdr_mapping)
+		free(obj->phdr_to_shdr_mapping);
 	if (obj->fd > 0)
 		close(obj->fd);
 	if (obj->mapped != NULL && obj->mapped != MAP_FAILED)

mld/seg.c

@@ -25,43 +25,29 @@ static int load_phdr(struct object *obj, struct segment *seg, size_t index)
  */
 static int load_shdr(struct object *obj, struct segment *seg, size_t index)
 {
-	bool found = false;
+	uint32_t shndx = obj->phdr_to_shdr_mapping[index];
+	Elf32_Shdr *hdr = &obj->shdr[shndx];
 
-	for (size_t i = 0; i < obj->shdr_len; i++) {
-		Elf32_Shdr *hdr = &obj->shdr[i];
-
-		// find shdr that matches  the offset in phdr
-		if (!PHDR_SHDR_MATCH(seg->phdr, hdr))
-			continue;
-
-		// get name
-		uint32_t name = B32(hdr->sh_name);
-		if (name >= obj->shstrtab->len) {
-			ERROR("section name index [%d] out of bounds", name);
-			return M_ERROR;
-		}
-		seg->name = &obj->shstrtab->data[name];
-
-		// map bytes
-		uint32_t len = B32(hdr->sh_size);
-		uint32_t off = B32(hdr->sh_offset);
-		seg->bytes = (unsigned char *) (obj->mapped + off);
-
-		if (BOUND_CHK(obj, len, off)) {
-			ERROR("cannot map seg %s:%d", seg->name, index);
-			return M_ERROR;
-		}
-
-		found = true;
-		seg->shdr = hdr;
-		seg->shdr_idx = i;
-		break;
-	}
-
-	if (!found) {
-		ERROR("cannot find shdr for segment [%d]", index);
+	// get name
+	uint32_t name = B32(hdr->sh_name);
+	if (name >= obj->shstrtab->len) {
+		ERROR("section name index [%d] out of bounds", name);
 		return M_ERROR;
 	}
+	seg->name = &obj->shstrtab->data[name];
+
+	// map bytes
+	uint32_t len = B32(hdr->sh_size);
+	uint32_t off = B32(hdr->sh_offset);
+	seg->bytes = (unsigned char *) (obj->mapped + off);
+
+	if (BOUND_CHK(obj, len, off)) {
+		ERROR("cannot map seg %s:%d", seg->name, index);
+		return M_ERROR;
+	}
+
+	seg->shdr = hdr;
+	seg->shdr_idx = shndx;
 
 	return M_SUCCESS;
 }

mld/symtab.c

@@ -56,6 +56,10 @@ int symtab_get(struct symbol_table *symtab, Elf32_Sym **res, const char *name,
 		if (strcmp(name, symname) != 0)
 			continue;
 
+		// ignore absolute symbols
+		if (B16(sym->st_shndx) == SHN_ABS)
+			continue;
+
 		// only allow retrevial of local variables from the
 		// same object
 		if (sym->st_info >> 4 != STB_GLOBAL && symtab->map != NULL &&

msim/ins.c

@@ -1,25 +1,74 @@
 #include <mips.h>
 #include <melf.h>
 #include <stdint.h>
-#include <stdio.h>
 #include <unistd.h>
-
+#include <merror.h>
 #include "sim.h"
 
-#define I16(n) ((int32_t)(uint16_t)(n))
+/* sign extension */
+#define SE(n) ((uint32_t)(int16_t)(n))
+/* sign extension 64bit */
+#define SE64(n) ((uint64_t)(int32_t)(n))
+/* signed sign extension 64bit */
+#define SSE64(n) ((int64_t)(int32_t)(n))
+/* shifted sign extention */
+#define SSE(n, s) (SE(n) << (s))
+/* zero extension */
+#define ZE(n) ((uint32_t)(uint16_t)(n))
+/* get vaddr from offset and base */
+#define VADDR(sim, ins) \
+	((sim->reg[ins.rs] /* base */) + \
+	 (SE(ins.offset) /* offset */))
+/* gets the low 32 bits of a 64 bit value */
+#define LO(n) (((1ULL << 32) - 1) & (n))
+/* gets the hi 32 bits of a 64 bit value */
+#define HI(n) ((n) >> 32)
+
+/* convert to a pointer */
+#define PTR(ptr, type) ((type *)(uintptr_t)(ptr))
+
+static void sim_delay_slot(struct simulator *sim)
+{
+	if (sim->args->jdelay == false)
+		return;
+
+	uint32_t ins = * (uint32_t *) (uintptr_t) sim->pc;
+	union mips_instruction_data data = { .raw = B32(ins) };
+	sim->pc += 4;
+
+	switch (data.op) {
+	case MIPS_OP_REGIMM:
+	case MIPS_OP_J:
+	case MIPS_OP_JAL:
+	case MIPS_OP_JALX:
+	case MIPS_OP_BEQ:
+	case MIPS_OP_BEQL:
+	case MIPS_OP_BNE:
+	case MIPS_OP_BNEL:
+	case MIPS_OP_BGTZ:
+	case MIPS_OP_BGTZL:
+	case MIPS_OP_BLEZ:
+	case MIPS_OP_BLEZL:
+		sim_dump(sim, "attempted to execute jump instruction in delay"
+			"slot (0b%05b)", data.op);
+	default:
+	}
+
+	sim_ins(sim, ins);
+}
 
 static void sim_ins_special_sop30(struct simulator *sim,
 				  union mips_instruction_data ins)
 {
 	switch (ins.shamt) {
 	case MIPS_SOP30_MUL:
-		sim->reg[ins.rd] = ((int64_t)sim->reg[ins.rs] /
-				(int64_t)sim->reg[ins.rt]) >> 0;
+		sim->reg[ins.rd] = (SSE64(sim->reg[ins.rs]) *
+				SSE64(sim->reg[ins.rt])) >> 0;
 		break;
 
 	case MIPS_SOP30_MUH:
-		sim->reg[ins.rd] = ((int64_t)sim->reg[ins.rs] /
-				(int64_t)sim->reg[ins.rt]) >> 32;
+		sim->reg[ins.rd] = (SSE64(sim->reg[ins.rs]) *
+				SSE64(sim->reg[ins.rt])) >> 32;
 		break;
 
 	default:
@@ -32,13 +81,13 @@ static void sim_ins_special_sop31(struct simulator *sim,
 {
 	switch (ins.shamt) {
 	case MIPS_SOP31_MULU:
-		sim->reg[ins.rd] = ((uint64_t)sim->reg[ins.rs] /
-				(uint64_t)sim->reg[ins.rt]) >> 0;
+		sim->reg[ins.rd] = (SE64(sim->reg[ins.rs]) *
+				SE64(sim->reg[ins.rt])) >> 0;
 		break;
 
 	case MIPS_SOP31_MUHU:
-		sim->reg[ins.rd] = ((uint64_t)sim->reg[ins.rs] /
-				(uint64_t)sim->reg[ins.rt]) >> 32;
+		sim->reg[ins.rd] = (SE64(sim->reg[ins.rs]) *
+				SE64(sim->reg[ins.rt])) >> 32;
 		break;
 
 	default:
@@ -51,13 +100,13 @@ static void sim_ins_special_sop32(struct simulator *sim,
 {
 	switch (ins.shamt) {
 	case MIPS_SOP32_DIV:
-		sim->reg[ins.rd] = (int32_t)sim->reg[ins.rs] /
-				(int32_t)sim->reg[ins.rt];
+		sim->reg[ins.rd] = (signed) sim->reg[ins.rs] /
+				(signed) sim->reg[ins.rt];
 		break;
 
 	case MIPS_SOP32_MOD:
-		sim->reg[ins.rd] = (int32_t)sim->reg[ins.rs] %
-				(int32_t)sim->reg[ins.rt];
+		sim->reg[ins.rd] = (signed) sim->reg[ins.rs] %
+				(signed) sim->reg[ins.rt];
 		break;
 
 	default:
@@ -87,8 +136,8 @@ static void sim_ins_special(struct simulator *sim,
 {
 	switch (ins.funct) {
 	case MIPS_FUNCT_ADD:
-		sim->reg[ins.rd] = (int32_t)sim->reg[ins.rs] +
-				(int32_t)sim->reg[ins.rt];
+		// TODO: trap on overflow
+		sim->reg[ins.rd] = sim->reg[ins.rs] + sim->reg[ins.rt];
 		break;
 
 	case MIPS_FUNCT_ADDU:
@@ -116,9 +165,10 @@ static void sim_ins_special(struct simulator *sim,
 		break;
 
 	case MIPS_FUNCT_JALR:
-		sim->reg[MIPS_REG_RA] = sim->pc;
+		sim->reg[ins.rd] = sim->pc + 4;
 		/* fall through */
 	case MIPS_FUNCT_JR:
+		sim_delay_slot(sim);
 		sim->pc = sim->reg[ins.rs];
 		break;
 
@@ -147,7 +197,8 @@ static void sim_ins_special(struct simulator *sim,
 		break;
 
 	case MIPS_FUNCT_SLT:
-		sim->reg[ins.rd] = (int32_t)sim->reg[ins.rs] < (int32_t)sim->reg[ins.rt] ? 1 : 0;
+		sim->reg[ins.rd] = (signed) sim->reg[ins.rs] <
+				(signed) sim->reg[ins.rt] ? 1 : 0;
 		break;
 
 	case MIPS_FUNCT_SLTU:
@@ -155,11 +206,12 @@ static void sim_ins_special(struct simulator *sim,
 		break;
 
 	case MIPS_FUNCT_SRA:
-		sim->reg[ins.rd] = (int32_t)sim->reg[ins.rt] >> ins.shamt;
+		sim->reg[ins.rd] = (signed) sim->reg[ins.rt] >> ins.shamt;
 		break;
 
 	case MIPS_FUNCT_SRAV:
-		sim->reg[ins.rd] = (int32_t)sim->reg[ins.rt] >> sim->reg[ins.rs];
+		sim->reg[ins.rd] = (signed) sim->reg[ins.rt] >>
+				sim->reg[ins.rs];
 		break;
 
 	case MIPS_FUNCT_SRL:
@@ -171,8 +223,8 @@ static void sim_ins_special(struct simulator *sim,
 		break;
 
 	case MIPS_FUNCT_SUB:
-		sim->reg[ins.rd] = (int32_t)sim->reg[ins.rs] -
-				(int32_t)sim->reg[ins.rt];
+		// TODO: trap on overflow
+		sim->reg[ins.rd] = sim->reg[ins.rs] - sim->reg[ins.rt];
 		break;
 
 	case MIPS_FUNCT_SUBU:
@@ -209,29 +261,33 @@ static void sim_ins_special(struct simulator *sim,
 static void sim_ins_regimm(struct simulator *sim,
 			   union mips_instruction_data ins)
 {
+	uint32_t pc = sim->pc;
+
 	switch (ins.bfunct) {
 	case MIPS_FUNCT_BGEZAL:
 	case MIPS_FUNCT_BGEZALL:
-		sim->reg[MIPS_REG_RA] = sim->pc;
+		sim->reg[MIPS_REG_RA] = sim->pc + 4;
 		/* fall through */
 	case MIPS_FUNCT_BGEZ:
 	case MIPS_FUNCT_BGEZL:
-		if ((int32_t)sim->reg[ins.rs] >= 0)
-			sim->pc += ins.offset << 2;
+		sim_delay_slot(sim);
+		if ((signed) sim->reg[ins.rs] >= 0)
+			sim->pc = pc + SSE(ins.offset, 2);
 		break;
 
 	case MIPS_FUNCT_BLTZAL:
 	case MIPS_FUNCT_BLTZALL:
-		sim->reg[MIPS_REG_RA] = sim->pc;
+		sim->reg[MIPS_REG_RA] = sim->pc + 4;
 		/* fall through */
 	case MIPS_FUNCT_BLTZ:
 	case MIPS_FUNCT_BLTZL:
-		if ((int32_t)sim->reg[ins.rs] < 0)
-			sim->pc += ins.offset << 2;
+		sim_delay_slot(sim);
+		if ((signed) sim->reg[ins.rs] < 0)
+			sim->pc = pc + SSE(ins.offset, 2);
 		break;
 
 	default:
-		sim_dump(sim, "unknown bfunct (0b%06b)", ins.bfunct);
+		sim_dump(sim, "unknown branch funct (0b%06b)", ins.bfunct);
 	}
 }
 
@@ -241,6 +297,7 @@ void sim_ins(struct simulator *sim, uint32_t raw)
 	union mips_instruction_data ins = {
 		.raw = B32(raw)
 	};
+	uint32_t pc = sim->pc;
 
 	// reset zero reg
 	sim->reg[MIPS_REG_ZERO] = 0;
@@ -256,73 +313,75 @@ void sim_ins(struct simulator *sim, uint32_t raw)
 
 	case MIPS_OP_ADDI:
 		sim->reg[ins.rt] = (int32_t)sim->reg[ins.rs] +
-				(int16_t) ins.immd;
+				SE(ins.immd);
 		break;
 
 	case MIPS_OP_ADDIU:
-		sim->reg[ins.rt] = sim->reg[ins.rs] + ins.immd;
+		sim->reg[ins.rt] = sim->reg[ins.rs] + SE(ins.immd);
 		break;
 
 	case MIPS_OP_ANDI:
-		sim->reg[ins.rt] = sim->reg[ins.rs] & ins.immd;
+		sim->reg[ins.rt] = sim->reg[ins.rs] & ZE(ins.immd);
 		break;
 
 	case MIPS_OP_BALC:
 		sim->reg[MIPS_REG_RA] = sim->pc;
 		/* fall through */
 	case MIPS_OP_BC:
-		sim->pc += ins.offs26 << 2;
+		sim->pc += SSE(ins.offs26, 2);
 		break;
 
 	case MIPS_OP_BEQ:
 	case MIPS_OP_BEQL:
+		sim_delay_slot(sim);
 		if (sim->reg[ins.rs] == sim->reg[ins.rt])
-			sim->pc += ins.offset << 2;
+			sim->pc = pc + SSE(ins.offset, 2);
 		break;
 
 	case MIPS_OP_BGTZ:
 	case MIPS_OP_BGTZL:
-		if ((int32_t)sim->reg[ins.rs] <= 0)
-			sim->pc += ins.offset << 2;
+		sim_delay_slot(sim);
+		if ((signed) sim->reg[ins.rs] <= 0)
+			sim->pc = pc + SSE(ins.offset, 2);
 		break;
 
 	case MIPS_OP_BLEZ:
 	case MIPS_OP_BLEZL:
-		if ((int32_t)sim->reg[ins.rs] <= 0)
-			sim->pc += ins.offset << 2;
+		sim_delay_slot(sim);
+		if ((signed) sim->reg[ins.rs] <= 0)
+			sim->pc = pc + SSE(ins.offset, 2);
 		break;
 
 	case MIPS_OP_BNE:
 	case MIPS_OP_BNEL:
+		sim_delay_slot(sim);
 		if (sim->reg[ins.rs] != sim->reg[ins.rt])
-			sim->pc += ins.offset << 2;
+			sim->pc = pc + SSE(ins.offset, 2);
 		break;
 
 	case MIPS_OP_JAL:
-		sim->reg[MIPS_REG_RA] = sim->pc;
+		sim->reg[MIPS_REG_RA] = sim->pc + 4;
 		/* fall through */
 	case MIPS_OP_J:
-		sim->pc = ins.target << 2;
+		sim_delay_slot(sim);
+		sim->pc &= 0xF0000000;
+		sim->pc |= ins.target << 2;
 		break;
 
 	case MIPS_OP_LB:
-		sim->reg[ins.rt] = * (int8_t *) (uintptr_t) (sim->reg[ins.rs]
-				+ ins.offset);
+		sim->reg[ins.rt] = *PTR(VADDR(sim, ins), int8_t);
 		break;
 
 	case MIPS_OP_LBU:
-		sim->reg[ins.rt] = * (uint8_t *) (uintptr_t) (sim->reg[ins.rs]
-				+ ins.offset);
+		sim->reg[ins.rt] = *PTR(VADDR(sim, ins), uint8_t);
 		break;
 
 	case MIPS_OP_LH:
-		sim->reg[ins.rt] = * (int16_t *) (uintptr_t) (sim->reg[ins.rs]
-				+ ins.offset);
+		sim->reg[ins.rt] = *PTR(VADDR(sim, ins), int16_t);
 		break;
 
 	case MIPS_OP_LHU:
-		sim->reg[ins.rt] = * (uint16_t *) (uintptr_t) (sim->reg[ins.rs]
-				+ ins.offset);
+		sim->reg[ins.rt] = *PTR(VADDR(sim, ins), uint16_t);
 		break;
 
 	case MIPS_OP_LUI:
@@ -330,32 +389,28 @@ void sim_ins(struct simulator *sim, uint32_t raw)
 		break;
 
 	case MIPS_OP_LW:
-		sim->reg[ins.rt] = * (uint32_t *) (uintptr_t) (sim->reg[ins.rs]
-				+ ins.offset);
+		sim->reg[ins.rt] = *PTR(VADDR(sim, ins), uint32_t);
 		break;
 
 	case MIPS_OP_SB:
-		* (uint8_t *) (uintptr_t) (sim->reg[ins.rs] +
-				+ ins.offset) = sim->reg[ins.rt];
+		*PTR(VADDR(sim, ins), uint8_t) = sim->reg[ins.rt];
 		break;
 
 	case MIPS_OP_SH:
-		* (uint16_t *) (uintptr_t) (sim->reg[ins.rs] +
-				ins.offset) = sim->reg[ins.rt];
+		*PTR(VADDR(sim, ins), uint16_t) = sim->reg[ins.rt];
 		break;
 
 	case MIPS_OP_SW:
-		* (uint32_t *) (uintptr_t) (sim->reg[ins.rs] +
-				ins.offset) = sim->reg[ins.rt];
+		*PTR(VADDR(sim, ins), uint32_t) = sim->reg[ins.rt];
 		break;
 
 	case MIPS_OP_SLTI:
-		sim->reg[ins.rt] = (int32_t)sim->reg[ins.rs] <
-				(int16_t)ins.immd ? 1 : 0;
+		sim->reg[ins.rt] = (signed) sim->reg[ins.rs] <
+				(signed) SE(ins.immd) ? 1 : 0;
 		break;
 
 	case MIPS_OP_SLTIU:
-		sim->reg[ins.rt] = sim->reg[ins.rs] < ins.immd ? 1 : 0;
+		sim->reg[ins.rt] = sim->reg[ins.rs] < SE(ins.immd) ? 1 : 0;
 		break;
 
 	case MIPS_OP_ORI:

msim/load.c

@@ -52,12 +52,15 @@ static int load_ehdr(struct simulator *sim, struct load_state *state)
 			  res |= assert_ehdr(&baseline.e_##name, \
 			  &ehdr.e_##name, size)                  \
 
+	// ignore abi ver
+	ehdr.e_ident[EI_ABIVERSION] = 0x00;
+
 	int res = 0;
 	EHDR_ASSERT(ident,     EI_NIDENT);
 	EHDR_ASSERT(type,      sizeof(Elf32_Half));
 	EHDR_ASSERT(machine,   sizeof(Elf32_Half));
 	EHDR_ASSERT(version,   sizeof(Elf32_Word));
-	EHDR_ASSERT(flags,     sizeof(Elf32_Word));
+	// EHDR_ASSERT(flags,     sizeof(Elf32_Word));
 	EHDR_ASSERT(ehsize,    sizeof(Elf32_Half));
 	EHDR_ASSERT(phentsize, sizeof(Elf32_Half));
 	EHDR_ASSERT(shentsize, sizeof(Elf32_Half));

msim/main.c

@@ -2,7 +2,6 @@
 #include <stdio.h>
 #include <unistd.h>
 #include <merror.h>
-#include <signal.h>
 
 #include "fault.h"
 #include "sim.h"
@@ -11,7 +10,7 @@ static void help(void) {
 	printf("usage: msim [options] executable\n\n");
 	printf("options:\n");
 	printf("\t-h\t\tprints this help messaage\n");
-	printf("\t-c\t\tdisable runtime memory checks (allows segfault)\n");
+	printf("\t-j\t\tdisable jump delay slot\n");
 	printf("\t-d\t\truns the debugger\n");
 }
 
@@ -21,18 +20,18 @@ int main(int argc, char **argv) {
 	struct simulator_args args = {
 		.executable = NULL,
 		.debug = false,
-		.memchk = true,
+		.jdelay = true,
 	};
 
 	int c;
 
-	while ((c = getopt(argc, argv, "hcd")) != 1) {
+	while ((c = getopt(argc, argv, "hjd")) != 1) {
 		switch (c) {
 		case 'h':
 			help();
 			return M_SUCCESS;
-		case 'c':
-			args.memchk = false;
+		case 'j':
+			args.jdelay = false;
 			break;
 		case 'd':
 			args.debug = true;

msim/sim.h

@@ -9,7 +9,7 @@
 /// arguments
 struct simulator_args {
 	char *executable;
-	bool memchk;
+	bool jdelay;
 	bool debug;
 };
 

test/masm/fncall.asm

@@ -11,6 +11,7 @@ result:
 	move	$t0, $a0
 	move	$v0, $t0
 	jr	$ra
+	nop
 
 main:
 	# save ra on stack
@@ -20,6 +21,7 @@ main:
 	# set return to 17
 	li	$a0, 17
 	jal	result
+	nop
 
 	# pop ra from stack
 	lw	$ra, 0($sp)
@@ -27,3 +29,4 @@ main:
 
 	# return result
 	jr	$ra
+	nop

test/masm/fpe.asm

@@ -15,3 +15,4 @@ main:
 	# return
 	li	$v0, 0
 	jr	$ra
+	nop

test/masm/hello.asm

@@ -24,3 +24,4 @@ main:
 	# return 1
 	li	$v0, 0
 	jr	$ra
+	nop

test/masm/recursion.asm

@@ -0,0 +1,41 @@
+# Copyright (c) 2024 Freya Murphy
+
+# file: recursion.asm
+# test: should recurse sum n..0 numbers
+
+.text
+.align 2
+.globl main
+
+main:
+	# init $a0
+	li	$a0, 5
+
+sum:
+	# save stack
+	addiu	$sp, $sp, -8
+	sw	$ra, 0($sp)
+	sw	$s0, 4($sp)
+
+	# load n from a0
+	move	$s0, $a0
+
+	# skip if n is zero
+	li	$v0, 0
+	beq	$s0, $zero, add
+	nop
+	jal	sum
+	addi	$a0, $s0, -1
+
+add:
+	# n = n + returned
+	add	$v0, $s0, $v0
+
+	# restore stack
+	lw	$ra, 0($sp)
+	lw	$s0, 4($sp)
+	addiu	$sp, $sp, 8
+
+	# return
+	jr	$ra
+	nop

test/masm/segv.asm

@@ -13,3 +13,4 @@ main:
 	# return
 	li	$v0, 0
 	jr	$ra
+	nop

test/out/fpe

@@ -12,4 +12,4 @@ $v1:	0x00000000 $t3:	0x00000000 $s3:	0x00000000 $k1:	0x00000000
 $a0:	0x00000000 $t4:	0x00000000 $s4:	0x00000000 $gp:	0x00000000
 $a1:	0x00000000 $t5:	0x00000000 $s5:	0x00000000 $sp:	0x10001000
 $a2:	0x00000000 $t6:	0x00000000 $s6:	0x00000000 $fp:	0x00000000
-$a3:	0x00000000 $t7:	0x00000000 $s7:	0x00000000 $ra:	0x0040001c
+$a3:	0x00000000 $t7:	0x00000000 $s7:	0x00000000 $ra:	0x00400024

test/out/recursion.status

@@ -0,0 +1 @@
+15

test/out/segv

@@ -12,4 +12,4 @@ $v1:	0x00000000 $t3:	0x00000000 $s3:	0x00000000 $k1:	0x00000000
 $a0:	0x00000000 $t4:	0x00000000 $s4:	0x00000000 $gp:	0x00000000
 $a1:	0x00000000 $t5:	0x00000000 $s5:	0x00000000 $sp:	0x10001000
 $a2:	0x00000000 $t6:	0x00000000 $s6:	0x00000000 $fp:	0x00000000
-$a3:	0x00000000 $t7:	0x00000000 $s7:	0x00000000 $ra:	0x00400018
+$a3:	0x00000000 $t7:	0x00000000 $s7:	0x00000000 $ra:	0x00400020

tools/test.sh

@@ -35,7 +35,10 @@ if [ $res = 0 ]; then
 	printf "\033[32mPASSED\033[0m\n"
 else
 	printf "\033[31mFAILED\033[0m\n"
-	diff <(echo $out) <(echo $rout)
+	if [ "$status" != "$rstatus" ]; then
+		printf "exit: $rstatus (should be $status)\n"
+	fi
+	diff -Nau <(printf "%s\n" "$out") <(printf "%s\n" "$rout")
 fi
 
 exit $res