1ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
2ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown.globl _start
3ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
4ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown_start:
5ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        # This code tests for the fldcw "load floating point command word"
6ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	#   instruction.  On most x86 processors the retired_instruction
7ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	#   performance counter counts this as one instruction.  However,
8ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	#   on Pentium 4 systems it counts as two.  Therefore this can
9ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	#   affect BBV results on such a system.
10ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	# fldcw is most often used to set the rouding mode when doing
11ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	#   floating point to integer conversions
12ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
13ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	# It is encoded as "d9 /5" which means
14ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	#   1101 1001 xx10 1yyy
15ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	# Where xx is the "mod" which will be 00, 01, or 10 indicating offset
16ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	#   and yyy is the register field
17ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
18ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
19ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
20ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        # these are instructions with similar encodings to fldcw
21ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	# that can cause false positives if the test isn't explicit enough
22ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brownsimilar:
23ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        fld1   	   	       		# d9 e8
24ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldl2t				# d9 e9
25ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldl2e				# d9 ea
26ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldpi				# d9 eb
27ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldlg2				# d9 ec
28ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldln2				# d9 ed
29ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldz				# d9 ee
30ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
31ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	# check some varied ways of calling fldcw
32ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
33ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
34ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	# offset on stack
35ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brownstack:
36ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	sub	$4,%esp			# allocate space on stack
37ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fnstcw	2(%esp)
38ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldcw	2(%esp)
39ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	add	$4,%esp			# restore stack
40ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
41ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	# 32-bit register
42ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
43ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fnstcw	cw
44ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	mov	$cw,%eax
45ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldcw	0(%eax)			# eax
46ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	mov	$cw,%ebx
47ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldcw	0(%ebx)			# ebx
48ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	mov	$cw,%ecx
49ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldcw	0(%ecx)			# ecx
50ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	mov	$cw,%edx
51ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldcw	0(%edx)			# edx
52ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
53ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	# register + 8-bit offset
54ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Browneight_bit:
55ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	mov	$cw,%eax
56ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	sub	$32,%eax
57ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
58ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldcw	32(%eax)		# eax + 8 bit offset
59ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	mov	%eax,%ebx
60ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldcw	32(%ebx)		# ebx + 8 bit offset
61ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	mov	%eax,%ecx
62ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldcw	32(%ecx)		# ecx + 8 bit offset
63ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	mov	%eax,%edx
64ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldcw	32(%edx)		# edx + 8 bit offset
65ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
66ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	# register + 32-bit offset
67ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brownthirtytwo_bit:
68ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	mov	$cw,%eax
69ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	sub	$30000,%eax
70ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
71ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldcw	30000(%eax)		# eax + 16 bit offset
72ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	mov	%eax,%ebx
73ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldcw	30000(%ebx)		# ebx + 16 bit offset
74ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	mov	%eax,%ecx
75ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldcw	30000(%ecx)		# ecx + 16 bit offset
76ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	mov	%eax,%edx
77ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldcw	30000(%edx)		# edx + 16 bit offset
78ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
79ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	# check an fp/integer conversion
80ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	# in a loop to give a bigger count
81ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
82ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	mov	$1024,%ecx
83ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brownbig_loop:
84ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
85ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldl	three			# load value onto fp stack
86ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fnstcw	saved_cw		# store control word to mem
87ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	movzwl	saved_cw, %eax		# load cw from mem, zero extending
88ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	movb	$12, %ah		# set cw for "round to zero"
89ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	movw	%ax, cw			# store back to memory
90ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldcw	cw   			# save new rounding mode
91ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fistpl	result			# save stack value as integer to mem
92ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	fldcw	saved_cw		# restore old cw
93ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
94ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	loop	big_loop		# loop to make the count more obvious
95ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
96ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	movl	result, %ebx		# sanity check to see if the
97ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	cmp	$3,%ebx			# result is the expected one
98ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	je	exit
99ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
100ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brownprint_error:
101ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	mov 	$4,%eax			# write syscall
102ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown#ifdef VGO_darwin
103ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	pushl	$1
104ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	pushl	$error
105ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	pushl	$22
106ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown#else
107ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	mov	$1,%ebx			# stdout
108ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	mov	$error,%ecx		# string
109ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	mov 	$22,%edx		# length of string
110ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown#endif
111ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	int 	$0x80
112ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
113ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brownexit:
114ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown#ifdef VGO_darwin
115ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	pushl	result
116ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown#else
117ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	movl	result, %ebx		# load converted value
118ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown#endif
119ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	movl	$1,	%eax		# SYSCALL_EXIT
120ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	int	$0x80
121ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
122ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
123ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
124ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown.data
125ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brownsaved_cw:	.long 0
126ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Browncw:  	.long	0
127ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brownresult: .long	0
128ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brownthree:	.long	0			# a floating point 3.0
129ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	.long	1074266112
130ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brownerror:	.asciz  "Error!  Wrong result!\n"
131