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