generating code for casts!

8 files changed, 419 insertions, 97 deletions

diff --git a/05/codegen.b b/05/codegen.b
index 1f0d077..2b0ec29 100644
--- a/05/codegen.b
+++ b/05/codegen.b
@@ -21,6 +21,7 @@ global codegen_second_pass ; = 0 on first global pass, 1 on second global pass
 global functions_addresses ; ident list of addresses
 global functions_labels ; ident list of ident lists of label addresses
 global curr_function_labels ; ident list of labels for current function (written to in 1st pass, read from in 2nd pass)
+global curr_function_return_type
 
 #define REG_RAX 0
 #define REG_RBX 3
@@ -76,7 +77,72 @@ function emit_mov_reg
 	*1code_output = n
 	code_output += 1
 	return
-	
+
+function emit_mov_rax_imm64
+	argument imm64
+	; 48 b8 IMM64
+	*2code_output = 0xb848
+	code_output += 2
+	*8code_output = imm64
+	code_output += 8
+	return
+
+function emit_movsx_rax_al
+	; 48 0f be c0
+	*4code_output = 0xc0be0f48
+	code_output += 4
+	return
+
+function emit_movsx_rax_ax
+	; 48 0f bf c0
+	*4code_output = 0xc0bf0f48
+	code_output += 4
+	return
+
+function emit_movsx_rax_eax
+	; 48 63 c0
+	*2code_output = 0x6348
+	code_output += 2
+	*1code_output = 0xc0
+	code_output += 1
+	return
+
+function emit_movzx_rax_al
+	; 48 0f b6 c0
+	*4code_output = 0xc0b60f48
+	code_output += 4
+	return
+
+function emit_movzx_rax_ax
+	; 48 0f b7 c0
+	*4code_output = 0xc0b70f48
+	code_output += 4
+	return
+
+function emit_mov_eax_eax
+	; 89 c0
+	*2code_output = 0xc089
+	code_output += 2
+	return
+
+function emit_mov_qword_rsp_plus_imm32_rax
+	argument imm32
+	; 48 89 84 24 IMM32
+	*4code_output = 0x24848948
+	code_output += 4
+	*4code_output = imm32
+	code_output += 4
+	return
+
+function emit_mov_rax_qword_rsp_plus_imm32
+	argument imm32
+	; 48 8b 84 24 IMM32
+	*4code_output = 0x24848b48
+	code_output += 4
+	*4code_output = imm32
+	code_output += 4
+	return
+
 
 function emit_sub_rsp_imm32
 	argument imm32
@@ -117,33 +183,6 @@ function emit_ret
 	code_output += 1
 	return
 
-function emit_mov_qword_rsp_plus_imm32_rax
-	argument imm32
-	; 48 89 84 24 IMM32
-	*4code_output = 0x24848948
-	code_output += 4
-	*4code_output = imm32
-	code_output += 4
-	return
-
-function emit_mov_rax_qword_rsp_plus_imm32
-	argument imm32
-	; 48 8b 84 24 IMM32
-	*4code_output = 0x24848b48
-	code_output += 4
-	*4code_output = imm32
-	code_output += 4
-	return
-
-function emit_mov_rax_imm64
-	argument imm64
-	; 48 b8 IMM64
-	*2code_output = 0xb848
-	code_output += 2
-	*8code_output = imm64
-	code_output += 8
-	return
-
 function emit_call_rax
 	; ff d0
 	*2code_output = 0xd0ff
@@ -185,6 +224,59 @@ function emit_movsq
 	code_output += 2
 	return
 
+function emit_movss_xmm0_dword_rax
+	; f3 0f 10 00
+	*4code_output = 0x00100ff3
+	code_output += 4
+	return
+
+function emit_movsd_xmm0_qword_rax
+	; f2 0f 10 00
+	*4code_output = 0x00100ff2
+	code_output += 4
+	return
+
+function emit_movss_dword_rax_xmm0
+	; f3 0f 11 00
+	*4code_output = 0x00110ff3
+	code_output += 4
+	return
+
+function emit_movsd_qword_rax_xmm0
+	; f2 0f 11 00
+	*4code_output = 0x00110ff2
+	code_output += 4
+	return
+
+function emit_cvtss2sd_xmm0_xmm0
+	; f3 0f 5a c0
+	*4code_output = 0xc05a0ff3
+	code_output += 4
+	return
+
+function emit_cvtsd2ss_xmm0_xmm0
+	; f2 0f 5a c0
+	*4code_output = 0xc05a0ff2
+	code_output += 4
+	return
+
+function emit_cvttsd2si_rax_xmm0
+	; f2 48 0f 2c c0
+	*4code_output = 0x2c0f48f2
+	code_output += 4
+	*1code_output = 0xc0
+	code_output += 1
+	return
+
+function emit_cvtsi2sd_xmm0_rax
+	; f2 48 0f 2a c0
+	*4code_output = 0x2a0f48f2
+	code_output += 4
+	*1code_output = 0xc0
+	code_output += 1
+	return
+
+
 ; make sure you put the return value in the proper place before calling this
 function generate_return
 	emit_mov_reg(REG_RSP, REG_RBP)
@@ -193,24 +285,6 @@ function generate_return
 	emit_ret()
 	return
 
-; returns pointer to end of expression
-function generate_push_expression
-	argument expr
-	local c
-	c = *1expr
-	if c == EXPRESSION_CONSTANT_INT goto generate_push_int
-	
-	die(.str_genpushexprNI)
-	:str_genpushexprNI
-		string generate_push_expression not implemented.
-		byte 0
-	:generate_push_int
-		expr += 8
-		emit_mov_rax_imm64(*8expr)
-		emit_push_rax()
-		expr += 8
-		return expr
-
 ; copy sizeof(type) bytes, rounded up to the nearest 8, from rsi to rdi
 function generate_copy_rsi_to_rdi_qwords
 	argument type
@@ -230,6 +304,225 @@ function generate_copy_rsi_to_rdi_qwords
 	emit_movsq()
 	return
 
+; cast whatever was just pushed onto the stack from from_type to to_type
+; `statement` is used for errors
+function generate_cast_top_of_stack
+	argument statement
+	argument from_type
+	argument to_type
+	local from
+	local to
+	local c
+	local d
+	
+	from = types + from_type
+	to = types + to_type
+	
+	if *1to == TYPE_VOID goto return_0 ; cast to void my ass
+	if *1from == TYPE_VOID goto bad_gen_cast ; cast from void to something - that's bad
+	if *1from == TYPE_ARRAY goto bad_gen_cast ; cast array (this probably won't ever happen because of decaying)
+	if *1to == TYPE_ARRAY goto bad_gen_cast ; cast to array
+	if *1from == TYPE_FUNCTION goto bad_gen_cast ; shouldn't happen
+	if *1to == TYPE_FUNCTION goto bad_gen_cast ; shouldn't happen
+	if *1to == TYPE_STRUCT goto gen_cast_to_struct
+	if *1from == TYPE_STRUCT goto bad_gen_cast ; cast from struct to something else
+	if *1to < TYPE_FLOAT goto gen_cast_to_integer
+	if *1to == TYPE_POINTER goto gen_cast_to_integer ; pointers are basically integers
+	
+	; cast to float/double
+	if *1from == TYPE_POINTER goto bad_gen_cast ; pointer to float/double
+	if *1to == *1from goto return_0
+	if *1from == TYPE_DOUBLE goto gen_cast_double_to_float
+	if *1from == TYPE_FLOAT goto gen_cast_float_to_double
+	; int to float/double
+	if *1to == TYPE_FLOAT goto gen_cast_int_to_float
+	if *1to == TYPE_DOUBLE goto gen_cast_int_to_double
+	
+	goto bad_gen_cast ; in theory we shouldn't get here
+	
+	:gen_cast_to_integer
+		if *1from == *1to goto return_0 ; casting from type to same type
+		if *1from == TYPE_POINTER goto return_0 ; no need to do anything
+		; cast float/double to integer
+		if *1from == TYPE_FLOAT goto gen_cast_float_to_int
+		if *1from == TYPE_DOUBLE goto gen_cast_double_to_int
+		
+		c = type_sizeof(*1from)
+		d = type_sizeof(*1to)
+		if d > c goto return_0 ; casting to bigger type, so we're good
+		if c == 8 goto return_0 ; casting from unsigned/signed long to unsigned/signed long, we're good
+		
+		; mov rax, [rsp]
+		emit_mov_rax_qword_rsp_plus_imm32(0)
+		
+		; now sign/zero extend the lower part of rax to the whole of rax		
+		if *1to == TYPE_CHAR goto gen_cast_integer_to_signed_char
+		if *1to == TYPE_UNSIGNED_CHAR goto gen_cast_integer_to_unsigned_char
+		if *1to == TYPE_SHORT goto gen_cast_integer_to_signed_short
+		if *1to == TYPE_UNSIGNED_SHORT goto gen_cast_integer_to_unsigned_short
+		if *1to == TYPE_INT goto gen_cast_integer_to_signed_int
+		if *1to == TYPE_UNSIGNED_INT goto gen_cast_integer_to_unsigned_int
+		
+		goto bad_gen_cast ; in theory we shouldn't get here
+		
+		:int2int_cast_cont
+		; mov [rsp], rax
+		emit_mov_qword_rsp_plus_imm32_rax(0)
+		return
+		
+		:gen_cast_integer_to_signed_char
+			emit_movsx_rax_al()
+			goto int2int_cast_cont
+		:gen_cast_integer_to_unsigned_char
+			emit_movzx_rax_al()
+			goto int2int_cast_cont
+		:gen_cast_integer_to_signed_short
+			emit_movsx_rax_ax()
+			goto int2int_cast_cont
+		:gen_cast_integer_to_unsigned_short
+			emit_movzx_rax_ax()
+			goto int2int_cast_cont
+		:gen_cast_integer_to_signed_int
+			emit_movsx_rax_eax()
+			goto int2int_cast_cont
+		:gen_cast_integer_to_unsigned_int
+			emit_mov_eax_eax()
+			goto int2int_cast_cont
+	:gen_cast_to_struct
+		; this is necessary because we add an implicit cast for return values
+		; so if we didn't have this, we wouldn't be able to return structs.
+		if *1from != TYPE_STRUCT goto bad_gen_cast
+		from += 1
+		to += 1
+		if *8from != *8to goto bad_gen_cast
+		return ; no casting needed; these are the same type
+	:gen_cast_double_to_float
+		; mov rax, rsp
+		emit_mov_reg(REG_RAX, REG_RSP)
+		; movsd xmm0, [rax]
+		emit_movsd_xmm0_qword_rax()
+		; cvtsd2ss xmm0, xmm0
+		emit_cvtsd2ss_xmm0_xmm0()
+		; movsd [rax], xmm0
+		emit_movsd_qword_rax_xmm0()
+		return
+	:gen_cast_float_to_double
+		; mov rax, rsp
+		emit_mov_reg(REG_RAX, REG_RSP)
+		; movss xmm0, [rax]
+		emit_movss_xmm0_dword_rax()
+		; cvtss2sd xmm0, xmm0
+		emit_cvtss2sd_xmm0_xmm0()
+		; movss [rax], xmm0
+		emit_movss_dword_rax_xmm0()
+		return
+	:gen_cast_int_to_float
+		; to reduce # of instructions, we first convert int to double, then double to float
+		; mov rax, [rsp]
+		emit_mov_rax_qword_rsp_plus_imm32(0)
+		; cvtsi2sd xmm0, rax
+		emit_cvtsi2sd_xmm0_rax()
+		; cvtsd2ss xmm0, xmm0
+		emit_cvtsd2ss_xmm0_xmm0()
+		; mov rax, rsp
+		emit_mov_reg(REG_RAX, REG_RSP)
+		; movss [rax], xmm0
+		emit_movss_dword_rax_xmm0()
+		; it shouldn't matter that there's junk at [rsp+4]
+		return
+	:gen_cast_int_to_double
+		; mov rax, [rsp]
+		emit_mov_rax_qword_rsp_plus_imm32(0)
+		; cvtsi2sd xmm0, rax
+		emit_cvtsi2sd_xmm0_rax()
+		; mov rax, rsp
+		emit_mov_reg(REG_RAX, REG_RSP)
+		; movsd [rax], xmm0
+		emit_movsd_qword_rax_xmm0()
+		return
+	:gen_cast_float_to_int
+		; mov rax, rsp
+		emit_mov_reg(REG_RAX, REG_RSP)
+		; movss xmm0, [rax]
+		emit_movss_xmm0_dword_rax()
+		; convert float to double, then double to int
+		; cvtss2sd xmm0, xmm0
+		emit_cvtss2sd_xmm0_xmm0()
+		; cvttsd2si rax, xmm0
+		emit_cvttsd2si_rax_xmm0()
+		; mov [rsp], rax
+		emit_mov_qword_rsp_plus_imm32_rax(0)
+		return
+	:gen_cast_double_to_int
+		; mov rax, rsp
+		emit_mov_reg(REG_RAX, REG_RSP)
+		; movsd xmm0, [rax]
+		emit_movsd_xmm0_qword_rax()
+		; cvttsd2si rax, xmm0
+		emit_cvttsd2si_rax_xmm0()
+		; mov [rsp], rax
+		emit_mov_qword_rsp_plus_imm32_rax(0)
+		return
+	
+	:bad_gen_cast
+		print_statement_location(statement)
+		puts(.str_bad_gen_cast1)
+		print_type(from_type)
+		puts(.str_bad_gen_cast2)
+		print_type(to_type)
+		putc(10)
+		exit(1)
+	:str_bad_gen_cast1
+		string : Error: Cannot convert type
+		byte 32
+		byte 0
+	:str_bad_gen_cast2
+		string  to type
+		byte 32
+		byte 0
+; `statement` is used for errors
+; returns pointer to end of expression
+function generate_push_expression
+	argument statement
+	argument expr
+	local b
+	local c
+	local type
+	type = expr + 4
+	type = *4type
+	
+	c = *1expr
+	if c == EXPRESSION_CONSTANT_INT goto generate_push_int
+	if c == EXPRESSION_CONSTANT_FLOAT goto generate_push_float
+	if c == EXPRESSION_CAST goto generate_cast
+	
+	die(.str_genpushexprNI)
+	:str_genpushexprNI
+		string generate_push_expression not implemented.
+		byte 0
+	:generate_cast
+		expr += 4
+		c = *4expr ; cast type
+		expr += 8
+		b = *4expr ; original type
+		expr -= 4
+		expr = generate_push_expression(statement, expr)
+		generate_cast_top_of_stack(statement, b, c)
+		return expr
+	:generate_push_float
+		expr += 8
+		emit_mov_rax_imm64(*8expr)
+		emit_push_rax()
+		generate_cast_top_of_stack(statement, TYPE_DOUBLE, type)
+		expr += 8
+		return expr
+	:generate_push_int
+		expr += 8
+		emit_mov_rax_imm64(*8expr)
+		emit_push_rax()
+		expr += 8
+		return expr
+
 function generate_statement
 	argument statement
 	local dat1
@@ -268,22 +561,29 @@ function generate_statement
 		return
 	:gen_return
 		if dat1 == 0 goto gen_return_noexpr
-		generate_push_expression(dat1)
+		generate_push_expression(statement, dat1)
+		p = dat1 + 4 ; pointer to dat1 type
+		generate_cast_top_of_stack(statement, *4p, curr_function_return_type)
 		; copy sizeof(return expression) rounded up to 8 bytes from [rsp] to [rbp+16]
 		emit_mov_reg(REG_RSI, REG_RSP)
 		emit_lea_rax_rbp_plus_imm32(16)
 		emit_mov_reg(REG_RDI, REG_RAX)
-		p = dat1 + 4
 		generate_copy_rsi_to_rdi_qwords(*4p)
 		
 		:gen_return_noexpr
 		generate_return()
 		return
+
 function generate_function
 	argument function_name
 	argument function_statement
+	local function_type
 	local out0
 	
+	function_type = ident_list_lookup(function_types, function_name)
+	
+	curr_function_return_type = functype_return_type(function_type)
+	
 	if codegen_second_pass != 0 goto genf_second_pass
 		curr_function_labels = ident_list_create(4000) ; ~ 200 labels per function should be plenty
 		ident_list_add(functions_labels, function_name, curr_function_labels)
@@ -334,8 +634,8 @@ function generate_functions
 	
 	:function_addr_mismatch
 		; address of function on 2nd pass doesn't line up with 1st pass
-		fputs(2, .str_function_addr_mismatch)
-		fputs(2, function_name)
+		puts(.str_function_addr_mismatch)
+		puts(function_name)
 		exit(1)
 	:str_function_addr_mismatch
 		string Function address on first pass doesn't match 2nd pass:
diff --git a/05/main.b b/05/main.b
index 0e356d7..7e570f5 100644
--- a/05/main.b
+++ b/05/main.b
@@ -108,27 +108,31 @@ function types_init
 	p += 1
 	
 	*8ptypes_bytes_used = p - types
-	return 
+	return
 
-function fprint_token_location
-	argument fd
+function print_token_location
 	argument token
 	token += 2
-	fprint_filename(fd, *2token)
+	print_filename(*2token)
 	token += 2
-	fputc(fd, ':)
-	fputn(fd, *4token)
+	putc(':)
+	putn(*4token)
 	return
 
+function print_statement_location
+	argument statement
+	; statements & tokens have the same format for locations!
+	print_token_location(statement)
+	return
+	
 ; accepts EITHER file index OR pointer to filename
-function fprint_filename
-	argument fd
+function print_filename
 	argument file
 	if file ] 65535 goto print_filename_string
 	file = file_get(file)
 	; (fallthrough)
 	:print_filename_string
-	fputs(2, file)
+	puts(file)
 	return
 
 ; accepts EITHER file index OR pointer to filename
@@ -136,12 +140,12 @@ function compile_error
 	argument file
 	argument line
 	argument message
-	fprint_filename(2, file)
-	fputc(2, ':)
-	fputn(2, line)
-	fputs(2, .str_error_prefix)
-	fputs(2, message)
-	fputc(2, 10)
+	print_filename(file)
+	putc(':)
+	putn(line)
+	puts(.str_error_prefix)
+	puts(message)
+	putc(10)
 	exit(1)
 
 function token_error
@@ -161,12 +165,12 @@ function compile_warning
 	argument file
 	argument line
 	argument message
-	fprint_filename(2, file)
-	fputc(2, ':)
-	fputn(2, line)
-	fputs(2, .str_warning_prefix)
-	fputs(2, message)
-	fputc(2, 10)
+	print_filename(file)
+	putc(':)
+	putn(line)
+	puts(.str_warning_prefix)
+	puts(message)
+	putc(10)
 	return
 	
 :str_error_prefix
@@ -306,7 +310,7 @@ function main
 	byte 0
 
 :mmap_output_fd_failed
-	fputs(2, .str_mmap_output_fd_failed)
+	puts(.str_mmap_output_fd_failed)
 	exit(1)
 :str_mmap_output_fd_failed
 	string Couldn't mmap output file.
@@ -314,7 +318,7 @@ function main
 	byte 0
 
 :usage_error
-	fputs(2, .str_usage_error)
+	puts(.str_usage_error)
 	exit(1)
 
 :str_usage_error
diff --git a/05/main.c b/05/main.c
index 5cbffbf..899e6b9 100644
--- a/05/main.c
+++ b/05/main.c
@@ -14,5 +14,5 @@ struct B { struct A *blah; }
 */
 
 int main(int argc, char **argv) {
-	return 42;
+	return (float)(long)"hello";
 }
diff --git a/05/parse.b b/05/parse.b
index 7f288ae..618045b 100644
--- a/05/parse.b
+++ b/05/parse.b
@@ -843,7 +843,9 @@ function parse_statement
 		if *1token == SYMBOL_SEMICOLON goto return_no_expr
 		n = token_next_semicolon_not_in_brackets(token)
 		*8out = expressions_end
+		p = expressions_end + 4 ; type of expression
 		expressions_end = parse_expression(token, n, expressions_end)
+		type_decay_array_to_pointer_in_place(*4p)
 		token = n + 16
 		:return_no_expr
 		out += 32
@@ -2400,7 +2402,7 @@ function type_length
 		type += type_length(type)
 		return type - start
 	:type_length_not_function
-	fputs(2, .str_type_length_bad_type)
+	puts(.str_type_length_bad_type)
 	exit(1)
 	:str_type_length_bad_type
 		string Bad type passed to type_length. This shouldn't happen.
@@ -2606,7 +2608,7 @@ function parse_expression
 	if c == EXPRESSION_DIV goto type_binary_usual
 	if c == EXPRESSION_REMAINDER goto type_binary_usual_integer
 	
-	fputs(2, .str_binop_this_shouldnt_happen)
+	puts(.str_binop_this_shouldnt_happen)
 	exit(1)
 	:str_binop_this_shouldnt_happen
 		string Bad binop symbol (this shouldn't happen).
@@ -2769,7 +2771,7 @@ function parse_expression
 		if c == EXPRESSION_DEREFERENCE goto unary_dereference
 		if c == EXPRESSION_PRE_INCREMENT goto unary_type_scalar_nopromote
 		if c == EXPRESSION_PRE_DECREMENT goto unary_type_scalar_nopromote
-		fputs(2, .str_unop_this_shouldnt_happen)
+		puts(.str_unop_this_shouldnt_happen)
 		exit(1)
 		:str_unop_this_shouldnt_happen
 			string Bad unary symbol (this shouldn't happen).
@@ -2798,7 +2800,9 @@ function parse_expression
 		*4out = a
 		out += 4
 		p += 16
+		a = out + 4 ; pointer to casted expression type
 		out = parse_expression(p, tokens_end, out)
+		type_decay_array_to_pointer_in_place(*4a)
 		return out
 		:bad_cast
 			token_error(tokens, .str_bad_cast)
@@ -2840,7 +2844,7 @@ function parse_expression
 		*4type = a
 		return out
 	:unary_bad_type
-		fprint_token_location(1, tokens)
+		print_token_location(tokens)
 		puts(.str_unary_bad_type)
 		print_type(a)
 		putc(10)
@@ -3244,7 +3248,7 @@ function type_sizeof
 	if c == TYPE_ARRAY goto sizeof_array
 	if c == TYPE_STRUCT goto sizeof_struct
 	
-	fputs(2, .str_sizeof_bad)
+	puts(.str_sizeof_bad)
 	exit(1)
 	:str_sizeof_bad
 		string type_sizeof bad type.
@@ -3309,7 +3313,7 @@ function type_alignof
 	if c == TYPE_ARRAY goto alignof_array
 	if c == TYPE_STRUCT goto alignof_struct
 	
-	fputs(2, .str_alignof_bad)
+	puts(.str_alignof_bad)
 	exit(1)
 	:str_alignof_bad
 		string type_alignof bad type.
@@ -3637,7 +3641,7 @@ function fit_to_type
 	if c == TYPE_LONG goto fit_to_type_long
 	if c == TYPE_UNSIGNED_LONG goto fit_to_type_ulong
 	if c == TYPE_POINTER goto fit_to_type_ulong
-	fputs(2, .str_bad_fit_to_type)
+	puts(.str_bad_fit_to_type)
 	exit(1)
 	:str_bad_fit_to_type
 		string Bad type passed to fit_to_type.
@@ -3726,7 +3730,7 @@ function bad_types_to_operator
 	argument type1
 	argument type2
 	
-	fprint_token_location(1, token)
+	print_token_location(token)
 	puts(.str_bad_types_to_operator)
 	print_type(type1)
 	puts(.str_space_and_space)
@@ -4276,7 +4280,7 @@ function print_type
 	if c == TYPE_ARRAY goto print_type_array
 	if c == TYPE_STRUCT goto print_type_struct
 	if c == TYPE_FUNCTION goto print_type_function
-	fputs(2, .str_bad_print_type)
+	puts(.str_bad_print_type)
 	putnln(type)
 	putnln(c)
 	putnln(types_bytes_used)
diff --git a/05/preprocess.b b/05/preprocess.b
index 31ab43e..9a7dd5c 100644
--- a/05/preprocess.b
+++ b/05/preprocess.b
@@ -603,10 +603,10 @@ function translation_phase_4
 		if b != 0 goto pp_directive_endif
 		goto unrecognized_directive
 	:pp_directive_error
-		fputs(2, filename)
-		fputc(2, ':)
-		fputn(2, line_number)
-		fputs(2, .str_directive_error)
+		puts(filename)
+		putc(':)
+		putn(line_number)
+		puts(.str_directive_error)
 		exit(1)
 	:str_directive_error
 		string : #error
diff --git a/05/tokenize.b b/05/tokenize.b
index 841f7df..809c9ff 100644
--- a/05/tokenize.b
+++ b/05/tokenize.b
@@ -15,7 +15,7 @@ function file_get
 	:file_got
 	return p
 	:file_uhoh
-	fputs(2, .str_bad_file_index)
+	puts(.str_bad_file_index)
 	exit(1)
 	:str_bad_file_index
 		string Bad file index. This shouldn't happen.
@@ -609,7 +609,7 @@ function print_tokens
 		if *1p == TOKEN_STRING_LITERAL goto print_token_string_literal
 		if *1p == TOKEN_IDENTIFIER goto print_token_identifier
 		if *1p == TOKEN_EOF goto print_token_eof
-		fputs(2, .str_print_bad_token)
+		puts(.str_print_bad_token)
 		exit(1)
 		:print_token_keyword
 			s = get_keyword_str(*1p)
diff --git a/05/util.b b/05/util.b
index 25ce686..42b6342 100644
--- a/05/util.b
+++ b/05/util.b
@@ -137,9 +137,9 @@ function max_signed
 	
 function file_error
 	argument name
-	fputs(2, .str_file_error)
-	fputs(2, name)
-	fputc(2, 10)
+	puts(.str_file_error)
+	puts(name)
+	putc(10)
 	exit(1)
 
 :str_file_error
@@ -149,8 +149,8 @@ function file_error
 
 function die
 	argument message
-	fputs(2, message)
-	fputc(2, 10)
+	puts(message)
+	putc(10)
 	exit(1)
 
 function ftruncate
@@ -162,7 +162,7 @@ function ftruncate
 	return
 	
 :ftruncate_failed
-	fputs(2, .str_ftruncate_failed)
+	puts(.str_ftruncate_failed)
 	exit(1)
 :str_ftruncate_failed
 	string ftruncated failed.
@@ -200,7 +200,7 @@ function malloc
 	return memory + 8
 
 :malloc_failed
-	fputs(2, .str_out_of_memory)
+	puts(.str_out_of_memory)
 	exit(1)
 	
 :str_out_of_memory
@@ -768,7 +768,7 @@ function leftmost_1bit
 	:leftmost1bit_found
 		return i
 	:leftmost1bit_0
-		fputs(2, .str_leftmost1bit_0)
+		puts(.str_leftmost1bit_0)
 		exit(1)
 	:str_leftmost1bit_0
 		string 0 passed to leftmost_1bit.
diff --git a/README.md b/README.md
index 12faae7..0643b3e 100644
--- a/README.md
+++ b/README.md
@@ -119,6 +119,12 @@ ax  bx  cx  dx  sp  bp  si  di
 │ mov r8, rax          │ 49 89 c0          │ set r8 to rax (only used for syscalls) │
 │ mov r9, rax          │ 49 89 c1          │ set r9 to rax (only used for syscalls) │
 │ mov r10, rax         │ 49 89 c2          │ set r10 to rax (only used for syscalls)│
+| movsx rax, al        | 48 0f be c0       | sign-extend al to rax                  |
+| movsx rax, ax        | 48 0f bf c0       | sign-extend ax to rax                  |
+| movsx rax, eax       | 48 63 c0          | sign-extend eax to rax                 |
+| movzx rax, al        | 48 0f b6 c0       | zero-extend al to rax                  |
+| movzx rax, ax        | 48 0f b7 c0       | zero-extend ax to rax                  |
+| mov eax, eax         | 89 c0             | zero-extend eax to rax                 |
 │ xchg rax, rbx        │ 48 93             │ exchange the values of rax and rbx     │
 │ mov [rbx], rax       │ 48 89 03          │ store rax as 8 bytes at address rbx    │
 │ mov rax, [rbx]       │ 48 8b 03          │ load 8 bytes from address rbx into rax │
@@ -175,6 +181,14 @@ ax  bx  cx  dx  sp  bp  si  di
 │ ja IMM32             │ 0f 87 IMM32       │ jump if "above" (like jg but unsigned) │
 │ jbe IMM32            │ 0f 86 IMM32       │ jump if below or equal to              │
 │ jae IMM32            │ 0f 83 IMM32       │ jump if above or equal to              │
+│ movss xmm0, [rax]    │ f3 0f 10 00       │ load the float at *rax into xmm0       │
+│ movsd xmm0, [rax]    │ f2 0f 10 00       │ load the double at *rax into xmm0      │
+│ movss [rax], xmm0    │ f3 0f 11 00       │ store the float in xmm0 at *rax        │
+│ movsd [rax], xmm0    │ f2 0f 11 00       │ store the double in xmm0 at *rax       │
+| cvtss2sd xmm0, xmm0  | f3 0f 5a c0       | convert xmm0 from float to double      |
+| cvtsd2ss xmm0, xmm0  | f2 0f 5a c0       | convert xmm0 from double to float      |
+| cvttsd2si rax, xmm0  | f2 48 0f 2c c0    | convert double in xmm0 to int in rax   |
+| cvtsi2sd xmm0, rax   | f2 48 0f 2a c0    | convert int in rax to double in xmm0   |
 │ call rax             │ ff d0             │ call the function at address rax       │
 │ ret                  │ c3                │ return from function                   │
 │ syscall              │ 0f 05             │ execute a system call                  │