[thirdparty/openssl.git] / crypto / ppccpuid.pl

#!/usr/bin/env perl

$flavour = shift;

$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
( $xlate="${dir}ppc-xlate.pl" and -f $xlate ) or
( $xlate="${dir}perlasm/ppc-xlate.pl" and -f $xlate) or
die "can't locate ppc-xlate.pl";

open STDOUT,"| $^X $xlate $flavour ".shift || die "can't call $xlate: $!";

if ($flavour=~/64/) {
    $CMPLI="cmpldi";
    $SHRLI="srdi";
    $SIGNX="extsw";
} else {
    $CMPLI="cmplwi";
    $SHRLI="srwi";
    $SIGNX="mr";
}

$code=<<___;
.machine	"any"
.text

.globl	.OPENSSL_fpu_probe
.align	4
.OPENSSL_fpu_probe:
	fmr	f0,f0
	blr
	.long	0
	.byte	0,12,0x14,0,0,0,0,0
.size	.OPENSSL_fpu_probe,.-.OPENSSL_fpu_probe
.globl	.OPENSSL_ppc64_probe
.align	4
.OPENSSL_ppc64_probe:
	fcfid	f1,f1
	extrdi	r0,r0,32,0
	blr
	.long	0
	.byte	0,12,0x14,0,0,0,0,0
.size	.OPENSSL_ppc64_probe,.-.OPENSSL_ppc64_probe

.globl	.OPENSSL_altivec_probe
.align	4
.OPENSSL_altivec_probe:
	.long	0x10000484	# vor	v0,v0,v0
	blr
	.long	0
	.byte	0,12,0x14,0,0,0,0,0
.size	.OPENSSL_altivec_probe,.-..OPENSSL_altivec_probe

.globl	.OPENSSL_crypto207_probe
.align	4
.OPENSSL_crypto207_probe:
	lvx_u	v0,0,r1
	vcipher	v0,v0,v0
	blr
	.long	0
	.byte	0,12,0x14,0,0,0,0,0
.size	.OPENSSL_crypto207_probe,.-.OPENSSL_crypto207_probe

.globl	.OPENSSL_wipe_cpu
.align	4
.OPENSSL_wipe_cpu:
	xor	r0,r0,r0
	fmr	f0,f31
	fmr	f1,f31
	fmr	f2,f31
	mr	r3,r1
	fmr	f3,f31
	xor	r4,r4,r4
	fmr	f4,f31
	xor	r5,r5,r5
	fmr	f5,f31
	xor	r6,r6,r6
	fmr	f6,f31
	xor	r7,r7,r7
	fmr	f7,f31
	xor	r8,r8,r8
	fmr	f8,f31
	xor	r9,r9,r9
	fmr	f9,f31
	xor	r10,r10,r10
	fmr	f10,f31
	xor	r11,r11,r11
	fmr	f11,f31
	xor	r12,r12,r12
	fmr	f12,f31
	fmr	f13,f31
	blr
	.long	0
	.byte	0,12,0x14,0,0,0,0,0
.size	.OPENSSL_wipe_cpu,.-.OPENSSL_wipe_cpu

.globl	.OPENSSL_atomic_add
.align	4
.OPENSSL_atomic_add:
Ladd:	lwarx	r5,0,r3
	add	r0,r4,r5
	stwcx.	r0,0,r3
	bne-	Ladd
	$SIGNX	r3,r0
	blr
	.long	0
	.byte	0,12,0x14,0,0,0,2,0
	.long	0
.size	.OPENSSL_atomic_add,.-.OPENSSL_atomic_add

.globl	.OPENSSL_rdtsc
.align	4
.OPENSSL_rdtsc:
___
$code.=<<___	if ($flavour =~ /64/);
	mftb	r3
___
$code.=<<___	if ($flavour !~ /64/);
Loop_rdtsc:
	mftbu	r5
	mftb	r3
	mftbu	r4
	cmplw	r4,r5
	bne	Loop_rdtsc
___
$code.=<<___;
	blr
	.long	0
	.byte	0,12,0x14,0,0,0,0,0
.size	.OPENSSL_rdtsc,.-.OPENSSL_rdtsc

.globl	.OPENSSL_cleanse
.align	4
.OPENSSL_cleanse:
	$CMPLI	r4,7
	li	r0,0
	bge	Lot
	$CMPLI	r4,0
	beqlr-
Little:	mtctr	r4
	stb	r0,0(r3)
	addi	r3,r3,1
	bdnz	\$-8
	blr
Lot:	andi.	r5,r3,3
	beq	Laligned
	stb	r0,0(r3)
	subi	r4,r4,1
	addi	r3,r3,1
	b	Lot
Laligned:
	$SHRLI	r5,r4,2
	mtctr	r5
	stw	r0,0(r3)
	addi	r3,r3,4
	bdnz	\$-8
	andi.	r4,r4,3
	bne	Little
	blr
	.long	0
	.byte	0,12,0x14,0,0,0,2,0
	.long	0
.size	.OPENSSL_cleanse,.-.OPENSSL_cleanse
___
{
my ($out,$cnt,$max)=("r3","r4","r5");
my ($tick,$lasttick)=("r6","r7");
my ($diff,$lastdiff)=("r8","r9");

$code.=<<___;
.globl	.OPENSSL_instrument_bus
.align	4
.OPENSSL_instrument_bus:
	mtctr	$cnt

	mftb	$lasttick		# collect 1st tick
	li	$diff,0

	dcbf	0,$out			# flush cache line
	lwarx	$tick,0,$out		# load and lock
	add	$tick,$tick,$diff
	stwcx.	$tick,0,$out
	stwx	$tick,0,$out

Loop:	mftb	$tick
	sub	$diff,$tick,$lasttick
	mr	$lasttick,$tick
	dcbf	0,$out			# flush cache line
	lwarx	$tick,0,$out		# load and lock
	add	$tick,$tick,$diff
	stwcx.	$tick,0,$out
	stwx	$tick,0,$out
	addi	$out,$out,4		# ++$out
	bdnz	Loop

	mr	r3,$cnt
	blr
	.long	0
	.byte	0,12,0x14,0,0,0,2,0
	.long	0
.size	.OPENSSL_instrument_bus,.-.OPENSSL_instrument_bus

.globl	.OPENSSL_instrument_bus2
.align	4
.OPENSSL_instrument_bus2:
	mr	r0,$cnt
	slwi	$cnt,$cnt,2

	mftb	$lasttick		# collect 1st tick
	li	$diff,0

	dcbf	0,$out			# flush cache line
	lwarx	$tick,0,$out		# load and lock
	add	$tick,$tick,$diff
	stwcx.	$tick,0,$out
	stwx	$tick,0,$out

	mftb	$tick			# collect 1st diff
	sub	$diff,$tick,$lasttick
	mr	$lasttick,$tick
	mr	$lastdiff,$diff
Loop2:
	dcbf	0,$out			# flush cache line
	lwarx	$tick,0,$out		# load and lock
	add	$tick,$tick,$diff
	stwcx.	$tick,0,$out
	stwx	$tick,0,$out

	addic.	$max,$max,-1
	beq	Ldone2

	mftb	$tick
	sub	$diff,$tick,$lasttick
	mr	$lasttick,$tick
	cmplw	7,$diff,$lastdiff
	mr	$lastdiff,$diff

	mfcr	$tick			# pull cr
	not	$tick,$tick		# flip bits
	rlwinm	$tick,$tick,1,29,29	# isolate flipped eq bit and scale

	sub.	$cnt,$cnt,$tick		# conditional --$cnt
	add	$out,$out,$tick		# conditional ++$out
	bne	Loop2

Ldone2:
	srwi	$cnt,$cnt,2
	sub	r3,r0,$cnt
	blr
	.long	0
	.byte	0,12,0x14,0,0,0,3,0
	.long	0
.size	.OPENSSL_instrument_bus2,.-.OPENSSL_instrument_bus2
___
}

$code =~ s/\`([^\`]*)\`/eval $1/gem;
print $code;
close STDOUT;
Commit	Line	Data
9c9c83cc AP	1	#!/usr/bin/env perl
9c9c83cc AP	2
addd641f	3	$flavour = shift;
9c9c83cc AP	4
	5	$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
	6	( $xlate="${dir}ppc-xlate.pl" and -f $xlate ) or
	7	( $xlate="${dir}perlasm/ppc-xlate.pl" and -f $xlate) or
	8	die "can't locate ppc-xlate.pl";
	9
addd641f	10	open STDOUT,"\| $^X $xlate $flavour ".shift \|\| die "can't call $xlate: $!";
9c9c83cc	11
addd641f	12	if ($flavour=~/64/) {
9c9c83cc AP	13	$CMPLI="cmpldi";
	14	$SHRLI="srdi";
	15	$SIGNX="extsw";
	16	} else {
	17	$CMPLI="cmplwi";
	18	$SHRLI="srwi";
	19	$SIGNX="mr";
	20	}
	21
	22	$code=<<___;
492279f6	23	.machine "any"
9c9c83cc AP	24	.text
9c9c83cc AP	25
81eae077 AP	26	.globl .OPENSSL_fpu_probe
	27	.align 4
	28	.OPENSSL_fpu_probe:
	29	fmr f0,f0
	30	blr
	31	.long 0
	32	.byte 0,12,0x14,0,0,0,0,0
	33	.size .OPENSSL_fpu_probe,.-.OPENSSL_fpu_probe
b4b48a10	34	.globl .OPENSSL_ppc64_probe
9c9c83cc	35	.align 4
b4b48a10 AP	36	.OPENSSL_ppc64_probe:
	37	fcfid f1,f1
	38	extrdi r0,r0,32,0
9c9c83cc	39	blr
67150340 AP	40	.long 0
67150340 AP	41	.byte 0,12,0x14,0,0,0,0,0
d6019e16	42	.size .OPENSSL_ppc64_probe,.-.OPENSSL_ppc64_probe
9c9c83cc	43
6415dd7b AP	44	.globl .OPENSSL_altivec_probe
	45	.align 4
	46	.OPENSSL_altivec_probe:
	47	.long 0x10000484 # vor v0,v0,v0
	48	blr
67150340 AP	49	.long 0
67150340 AP	50	.byte 0,12,0x14,0,0,0,0,0
d6019e16	51	.size .OPENSSL_altivec_probe,.-..OPENSSL_altivec_probe
6415dd7b	52
de51e830 AP	53	.globl .OPENSSL_crypto207_probe
de51e830 AP	54	.align 4
d86689e1	55	.OPENSSL_crypto207_probe:
de51e830 AP	56	lvx_u v0,0,r1
	57	vcipher v0,v0,v0
	58	blr
	59	.long 0
	60	.byte 0,12,0x14,0,0,0,0,0
	61	.size .OPENSSL_crypto207_probe,.-.OPENSSL_crypto207_probe
	62
9c9c83cc AP	63	.globl .OPENSSL_wipe_cpu
	64	.align 4
	65	.OPENSSL_wipe_cpu:
	66	xor r0,r0,r0
78a533cb AP	67	fmr f0,f31
	68	fmr f1,f31
	69	fmr f2,f31
9c9c83cc	70	mr r3,r1
78a533cb	71	fmr f3,f31
9c9c83cc	72	xor r4,r4,r4
78a533cb	73	fmr f4,f31
9c9c83cc	74	xor r5,r5,r5
78a533cb	75	fmr f5,f31
9c9c83cc	76	xor r6,r6,r6
78a533cb	77	fmr f6,f31
9c9c83cc	78	xor r7,r7,r7
78a533cb	79	fmr f7,f31
9c9c83cc	80	xor r8,r8,r8
78a533cb	81	fmr f8,f31
9c9c83cc	82	xor r9,r9,r9
78a533cb	83	fmr f9,f31
9c9c83cc	84	xor r10,r10,r10
78a533cb	85	fmr f10,f31
9c9c83cc	86	xor r11,r11,r11
78a533cb	87	fmr f11,f31
9c9c83cc	88	xor r12,r12,r12
78a533cb AP	89	fmr f12,f31
78a533cb AP	90	fmr f13,f31
9c9c83cc	91	blr
67150340 AP	92	.long 0
67150340 AP	93	.byte 0,12,0x14,0,0,0,0,0
d6019e16	94	.size .OPENSSL_wipe_cpu,.-.OPENSSL_wipe_cpu
9c9c83cc AP	95
	96	.globl .OPENSSL_atomic_add
	97	.align 4
	98	.OPENSSL_atomic_add:
5fabb88a	99	Ladd: lwarx r5,0,r3
9c9c83cc AP	100	add r0,r4,r5
9c9c83cc AP	101	stwcx. r0,0,r3
5fabb88a	102	bne- Ladd
9c9c83cc AP	103	$SIGNX r3,r0
9c9c83cc AP	104	blr
67150340 AP	105	.long 0
	106	.byte 0,12,0x14,0,0,0,2,0
	107	.long 0
d6019e16	108	.size .OPENSSL_atomic_add,.-.OPENSSL_atomic_add
9c9c83cc AP	109
	110	.globl .OPENSSL_rdtsc
	111	.align 4
	112	.OPENSSL_rdtsc:
81eae077 AP	113	___
	114	$code.=<<___ if ($flavour =~ /64/);
	115	mftb r3
	116	___
	117	$code.=<<___ if ($flavour !~ /64/);
	118	Loop_rdtsc:
	119	mftbu r5
9c9c83cc AP	120	mftb r3
9c9c83cc AP	121	mftbu r4
81eae077 AP	122	cmplw r4,r5
	123	bne Loop_rdtsc
	124	___
	125	$code.=<<___;
9c9c83cc	126	blr
67150340 AP	127	.long 0
67150340 AP	128	.byte 0,12,0x14,0,0,0,0,0
d6019e16	129	.size .OPENSSL_rdtsc,.-.OPENSSL_rdtsc
9c9c83cc AP	130
	131	.globl .OPENSSL_cleanse
	132	.align 4
	133	.OPENSSL_cleanse:
	134	$CMPLI r4,7
	135	li r0,0
	136	bge Lot
7676eebf AP	137	$CMPLI r4,0
7676eebf AP	138	beqlr-
9c9c83cc AP	139	Little: mtctr r4
	140	stb r0,0(r3)
	141	addi r3,r3,1
9474483a	142	bdnz \$-8
9c9c83cc AP	143	blr
	144	Lot: andi. r5,r3,3
	145	beq Laligned
	146	stb r0,0(r3)
	147	subi r4,r4,1
	148	addi r3,r3,1
	149	b Lot
	150	Laligned:
	151	$SHRLI r5,r4,2
	152	mtctr r5
	153	stw r0,0(r3)
	154	addi r3,r3,4
9474483a	155	bdnz \$-8
9c9c83cc AP	156	andi. r4,r4,3
	157	bne Little
	158	blr
67150340 AP	159	.long 0
	160	.byte 0,12,0x14,0,0,0,2,0
	161	.long 0
d6019e16	162	.size .OPENSSL_cleanse,.-.OPENSSL_cleanse
9c9c83cc	163	___
5fabb88a AP	164	{
	165	my ($out,$cnt,$max)=("r3","r4","r5");
	166	my ($tick,$lasttick)=("r6","r7");
	167	my ($diff,$lastdiff)=("r8","r9");
	168
	169	$code.=<<___;
	170	.globl .OPENSSL_instrument_bus
	171	.align 4
	172	.OPENSSL_instrument_bus:
	173	mtctr $cnt
	174
	175	mftb $lasttick # collect 1st tick
	176	li $diff,0
	177
	178	dcbf 0,$out # flush cache line
	179	lwarx $tick,0,$out # load and lock
	180	add $tick,$tick,$diff
	181	stwcx. $tick,0,$out
	182	stwx $tick,0,$out
	183
	184	Loop: mftb $tick
	185	sub $diff,$tick,$lasttick
	186	mr $lasttick,$tick
	187	dcbf 0,$out # flush cache line
	188	lwarx $tick,0,$out # load and lock
	189	add $tick,$tick,$diff
	190	stwcx. $tick,0,$out
	191	stwx $tick,0,$out
	192	addi $out,$out,4 # ++$out
	193	bdnz Loop
	194
	195	mr r3,$cnt
	196	blr
67150340 AP	197	.long 0
	198	.byte 0,12,0x14,0,0,0,2,0
	199	.long 0
d6019e16	200	.size .OPENSSL_instrument_bus,.-.OPENSSL_instrument_bus
5fabb88a AP	201
	202	.globl .OPENSSL_instrument_bus2
	203	.align 4
	204	.OPENSSL_instrument_bus2:
	205	mr r0,$cnt
	206	slwi $cnt,$cnt,2
	207
	208	mftb $lasttick # collect 1st tick
	209	li $diff,0
	210
	211	dcbf 0,$out # flush cache line
	212	lwarx $tick,0,$out # load and lock
	213	add $tick,$tick,$diff
	214	stwcx. $tick,0,$out
	215	stwx $tick,0,$out
	216
	217	mftb $tick # collect 1st diff
	218	sub $diff,$tick,$lasttick
	219	mr $lasttick,$tick
	220	mr $lastdiff,$diff
	221	Loop2:
	222	dcbf 0,$out # flush cache line
	223	lwarx $tick,0,$out # load and lock
	224	add $tick,$tick,$diff
	225	stwcx. $tick,0,$out
	226	stwx $tick,0,$out
	227
	228	addic. $max,$max,-1
	229	beq Ldone2
	230
	231	mftb $tick
	232	sub $diff,$tick,$lasttick
	233	mr $lasttick,$tick
	234	cmplw 7,$diff,$lastdiff
	235	mr $lastdiff,$diff
	236
	237	mfcr $tick # pull cr
	238	not $tick,$tick # flip bits
	239	rlwinm $tick,$tick,1,29,29 # isolate flipped eq bit and scale
	240
	241	sub. $cnt,$cnt,$tick # conditional --$cnt
	242	add $out,$out,$tick # conditional ++$out
	243	bne Loop2
	244
	245	Ldone2:
	246	srwi $cnt,$cnt,2
	247	sub r3,r0,$cnt
	248	blr
67150340 AP	249	.long 0
	250	.byte 0,12,0x14,0,0,0,3,0
	251	.long 0
d6019e16	252	.size .OPENSSL_instrument_bus2,.-.OPENSSL_instrument_bus2
5fabb88a AP	253	___
5fabb88a AP	254	}
9c9c83cc AP	255
	256	$code =~ s/\`([^\`]*)\`/eval $1/gem;
	257	print $code;
	258	close STDOUT;