AES for IA64 update [from HEAD].

1 files changed, 645 insertions, 1174 deletions

diff --git a/crypto/aes/asm/aes-ia64.S b/crypto/aes/asm/aes-ia64.S
index 542cf335e9..7f6c4c3662 100644
--- a/crypto/aes/asm/aes-ia64.S
+++ b/crypto/aes/asm/aes-ia64.S
@@ -17,14 +17,24 @@
 // big-endian input, ECB timing on Itanium 2 is (18 + 13*rounds)
 // ticks per block, or 9.25 CPU cycles per byte for 128 bit key.
 
-.ident	"aes-ia64.S, version 1.1"
+// Version 1.2 mitigates the hazard of cache-timing attacks by
+// a) compressing S-boxes from 8KB to 2KB+256B, b) scheduling
+// references to S-boxes for L2 cache latency, c) prefetching T[ed]4
+// prior last round. As result performance dropped to (26 + 15*rounds)
+// ticks per block or 11 cycles per byte processed with 128-bit key.
+// This is ~16% deterioration. For reference Itanium 2 L1 cache has
+// 64 bytes line size and L2 - 128 bytes...
+
+.ident	"aes-ia64.S, version 1.2"
 .ident	"IA-64 ISA artwork by Andy Polyakov <appro@fy.chalmers.se>"
 .explicit
 .text
 
 rk0=r8;     rk1=r9;
 
-prsave=r10;
+pfssave=r2;
+lcsave=r10;
+prsave=r3;
 maskff=r11;
 twenty4=r14;
 sixteen=r15;
@@ -44,12 +54,21 @@ te0=r40;    te1=r41;    te2=r42;    te3=r43;
 
 #if defined(_HPUX_SOURCE) && !defined(_LP64)
 # define ADDP	addp4
-# define KSZ	4
-# define LDKEY	ld4
 #else
 # define ADDP	add
 #endif
 
+// Offsets from Te0
+#define TE0	0
+#define TE2	2
+#if defined(_HPUX_SOURCE) || defined(B_ENDIAN)
+#define TE1	3
+#define TE3	1
+#else
+#define TE1	1
+#define TE3	3
+#endif
+
 // This implies that AES_KEY comprises 32-bit key schedule elements
 // even on LP64 platforms.
 #ifndef	KSZ
@@ -67,16 +86,19 @@ te0=r40;    te1=r41;    te2=r42;    te3=r43;
 // Clobber:	r16-r31,rk0-rk1,r32-r43
 .align	32
 _ia64_AES_encrypt:
+	.prologue
+	.altrp	b6
+	.body
 { .mmi;	alloc	r16=ar.pfs,12,0,0,8
 	LDKEY	t0=[rk0],2*KSZ
 	mov	pr.rot=1<<16	}
 { .mmi;	LDKEY	t1=[rk1],2*KSZ
-	add	te1=1024,te0
+	add	te1=TE1,te0
 	add	te3=-3,te3	};;
 { .mib;	LDKEY	t2=[rk0],2*KSZ
-	mov	ar.ec=3		}
+	mov	ar.ec=2		}
 { .mib;	LDKEY	t3=[rk1],2*KSZ
-	add	te2=2048,te0
+	add	te2=TE2,te0
 	brp.loop.imp	.Le_top,.Le_end-16	};;
 
 { .mmi;	xor	s0=s0,t0
@@ -84,8 +106,8 @@ _ia64_AES_encrypt:
 	mov	ar.lc=te3	}
 { .mmi;	xor	s2=s2,t2
 	xor	s3=s3,t3
-	add	te3=3072,te0	};;
-
+	add	te3=TE3,te0	};;
+
 .align	32
 .Le_top:
 { .mmi;	(p0)	LDKEY	t0=[rk0],2*KSZ		// 0/0:rk[0]
@@ -95,105 +117,184 @@ _ia64_AES_encrypt:
 	(p0)	and	te30=s0,maskff		// 0/1:s0&0xff
 	(p0)	shr.u	te00=s0,twenty4	};;	// 0/0:s0>>24
 { .mmi;	(p0)	LDKEY	t2=[rk0],2*KSZ		// 1/2:rk[2]
-	(p0)	shladd	te33=te33,2,te3		// 1/0:te0+s0>>24
+	(p0)	shladd	te33=te33,3,te3		// 1/0:te0+s0>>24
 	(p0)	extr.u	te23=s3,8,8	}	// 1/1:s3>>8&0xff
 { .mmi;	(p0)	LDKEY	t3=[rk1],2*KSZ		// 1/3:rk[3]
-	(p0)	shladd	te30=te30,2,te3		// 1/1:te3+s0
+	(p0)	shladd	te30=te30,3,te3		// 1/1:te3+s0
 	(p0)	shr.u	te01=s1,twenty4	};;	// 1/1:s1>>24
 { .mmi;	(p0)	ld4	te33=[te33]		// 2/0:te3[s3&0xff]
-	(p0)	shladd	te22=te22,2,te2		// 2/0:te2+s2>>8&0xff
+	(p0)	shladd	te22=te22,3,te2		// 2/0:te2+s2>>8&0xff
 	(p0)	extr.u	te20=s0,8,8	}	// 2/2:s0>>8&0xff
 { .mmi;	(p0)	ld4	te30=[te30]		// 2/1:te3[s0]
-	(p0)	shladd	te23=te23,2,te2		// 2/1:te2+s3>>8
+	(p0)	shladd	te23=te23,3,te2		// 2/1:te2+s3>>8
 	(p0)	shr.u	te02=s2,twenty4	};;	// 2/2:s2>>24
 { .mmi;	(p0)	ld4	te22=[te22]		// 3/0:te2[s2>>8]
-	(p0)	shladd	te20=te20,2,te2		// 3/2:te2+s0>>8
+	(p0)	shladd	te20=te20,3,te2		// 3/2:te2+s0>>8
 	(p0)	extr.u	te21=s1,8,8	}	// 3/3:s1>>8&0xff
 { .mmi;	(p0)	ld4	te23=[te23]		// 3/1:te2[s3>>8]
-	(p0)	shladd	te00=te00,2,te0		// 3/0:te0+s0>>24
+	(p0)	shladd	te00=te00,3,te0		// 3/0:te0+s0>>24
 	(p0)	shr.u	te03=s3,twenty4	};;	// 3/3:s3>>24
 { .mmi;	(p0)	ld4	te20=[te20]		// 4/2:te2[s0>>8]
-	(p0)	shladd	te21=te21,2,te2		// 4/3:te3+s2
+	(p0)	shladd	te21=te21,3,te2		// 4/3:te3+s2
 	(p0)	extr.u	te11=s1,16,8	}	// 4/0:s1>>16&0xff
 { .mmi;	(p0)	ld4	te00=[te00]		// 4/0:te0[s0>>24]
-	(p0)	shladd	te01=te01,2,te0		// 4/1:te0+s1>>24
+	(p0)	shladd	te01=te01,3,te0		// 4/1:te0+s1>>24
 	(p0)	shr.u	te13=s3,sixteen	};;	// 4/2:s3>>16
 { .mmi;	(p0)	ld4	te21=[te21]		// 5/3:te2[s1>>8]
-	(p0)	shladd	te11=te11,2,te1		// 5/0:te1+s1>>16
+	(p0)	shladd	te11=te11,3,te1		// 5/0:te1+s1>>16
 	(p0)	extr.u	te12=s2,16,8	}	// 5/1:s2>>16&0xff
 { .mmi;	(p0)	ld4	te01=[te01]		// 5/1:te0[s1>>24]
-	(p0)	shladd	te02=te02,2,te0		// 5/2:te0+s2>>24
+	(p0)	shladd	te02=te02,3,te0		// 5/2:te0+s2>>24
 	(p0)	and	te31=s1,maskff	};;	// 5/2:s1&0xff
-
 { .mmi;	(p0)	ld4	te11=[te11]		// 6/0:te1[s1>>16]
-	(p0)	shladd	te12=te12,2,te1		// 6/1:te1+s2>>16
+	(p0)	shladd	te12=te12,3,te1		// 6/1:te1+s2>>16
 	(p0)	extr.u	te10=s0,16,8	}	// 6/3:s0>>16&0xff
 { .mmi;	(p0)	ld4	te02=[te02]		// 6/2:te0[s2>>24]
-	(p0)	shladd	te03=te03,2,te0		// 6/3:te1+s0>>16
+	(p0)	shladd	te03=te03,3,te0		// 6/3:te1+s0>>16
 	(p0)	and	te32=s2,maskff	};;	// 6/3:s2&0xff
+
 { .mmi;	(p0)	ld4	te12=[te12]		// 7/1:te1[s2>>16]
-	(p0)	shladd	te31=te31,2,te3		// 7/2:te3+s1&0xff
+	(p0)	shladd	te31=te31,3,te3		// 7/2:te3+s1&0xff
 	(p0)	and	te13=te13,maskff}	// 7/2:s3>>16&0xff
 { .mmi;	(p0)	ld4	te03=[te03]		// 7/3:te0[s3>>24]
-	(p0)	shladd	te32=te32,2,te3		// 7/3:te3+s2
+	(p0)	shladd	te32=te32,3,te3		// 7/3:te3+s2
 	(p0)	xor	t0=t0,te33	};;	// 7/0:
 { .mmi;	(p0)	ld4	te31=[te31]		// 8/2:te3[s1]
-	(p0)	shladd	te13=te13,2,te1		// 8/2:te1+s3>>16
+	(p0)	shladd	te13=te13,3,te1		// 8/2:te1+s3>>16
 	(p0)	xor	t0=t0,te22	}	// 8/0:
 { .mmi;	(p0)	ld4	te32=[te32]		// 8/3:te3[s2]
-	(p0)	shladd	te10=te10,2,te1		// 8/3:te1+s0>>16
+	(p0)	shladd	te10=te10,3,te1		// 8/3:te1+s0>>16
 	(p0)	xor	t1=t1,te30	};;	// 8/1:
 { .mmi;	(p0)	ld4	te13=[te13]		// 9/2:te1[s3>>16]
-	(p0)	xor	t0=t0,te00		// 9/0:
-	(p0)	xor	t1=t1,te23	}	// 9/1:		
-{ .mmi;	(p0)	ld4	te10=[te10]		// 9/3:te1[s0>>16]
-	(p0)	xor	t2=t2,te20		// 9/2:
-	(p0)	xor	t3=t3,te21	};;	// 9/3:
-{ .mmi;	(p0)	xor	t0=t0,te11		// 10/0:done!
-	(p0)	xor	t1=t1,te01		// 10/1:
-	(p0)	xor	t2=t2,te02	}	// 10/2:
-{ .mmi;	(p0)	xor	t3=t3,te03		// 10/3:
-	(p16)	cmp.eq	p0,p17=r0,r0 	};;	// 10/clear (p17)
-{ .mmi;	(p0)	xor	t1=t1,te12		// 11/1:done!
-	(p0)	xor	t2=t2,te31		// 11/2:
-	(p0)	xor	t3=t3,te32	}	// 11/3:
-{ .mmi;	(p17)	add	te0=4096,te0		// 11/	
-	(p17)	add	te1=4096,te1	};;	// 11/
-{ .mib;	(p0)	xor	t2=t2,te13		// 12/2:done!
-	(p0)	xor	t3=t3,te10	}	// 12/3:done!
-{ .mib;	(p17)	add	te2=4096,te2		// 12/
-	(p17)	add	te3=4096,te3		// 12/
+	(p0)	ld4	te10=[te10]		// 9/3:te1[s0>>16]
+	(p0)	xor	t0=t0,te00	};;	// 9/0:		!L2 scheduling
+{ .mmi;	(p0)	xor	t1=t1,te23		// 10[9]/1:	
+	(p0)	xor	t2=t2,te20		// 10[9]/2:
+	(p0)	xor	t3=t3,te21	};;	// 10[9]/3:
+{ .mmi;	(p0)	xor	t0=t0,te11		// 11[10]/0:done!
+	(p0)	xor	t1=t1,te01		// 11[10]/1:
+	(p0)	xor	t2=t2,te02	};;	// 11[10]/2:	!L2 scheduling
+{ .mmi;	(p0)	xor	t3=t3,te03		// 12[10]/3:
+	(p16)	cmp.eq	p0,p17=r0,r0 	};;	// 12[10]/clear (p17)
+{ .mmi;	(p0)	xor	t1=t1,te12		// 13[11]/1:done!
+	(p0)	xor	t2=t2,te31		// 13[11]/2:
+	(p0)	xor	t3=t3,te32	}	// 13[11]/3:
+{ .mmi;	(p17)	add	te0=2048,te0		// 13[11]/
+	(p17)	add	te1=2048+64-TE1,te1};;	// 13[11]/
+{ .mib;	(p0)	xor	t2=t2,te13		// 14[12]/2:done!
+	(p17)	add	te2=2048+128-TE2,te2}	// 14[12]/
+{ .mib;	(p0)	xor	t3=t3,te10		// 14[12]/3:done!
+	(p17)	add	te3=2048+192-TE3,te3	// 14[12]/
 	br.ctop.sptk	.Le_top		};;
 .Le_end:
-{ .mib;	mov	r16=s0
-	mov	r20=s1			}
-{ .mib;	mov	r24=s2
-	mov	r28=s3
-	br.ret.sptk	b6		};;
+
+
+{ .mmi;	ld8	te12=[te0]		// prefetch Te4
+	ld8	te31=[te1]	}
+{ .mmi;	ld8	te10=[te2]
+	ld8	te32=[te3]	}
+
+{ .mmi;	LDKEY	t0=[rk0],2*KSZ		// 0/0:rk[0]
+	and	te33=s3,maskff		// 0/0:s3&0xff
+	extr.u	te22=s2,8,8	}	// 0/0:s2>>8&0xff
+{ .mmi; LDKEY	t1=[rk1],2*KSZ		// 0/1:rk[1]
+	and	te30=s0,maskff		// 0/1:s0&0xff
+	shr.u	te00=s0,twenty4	};;	// 0/0:s0>>24
+{ .mmi;	LDKEY	t2=[rk0],2*KSZ		// 1/2:rk[2]
+	add	te33=te33,te0		// 1/0:te0+s0>>24
+	extr.u	te23=s3,8,8	}	// 1/1:s3>>8&0xff
+{ .mmi;	LDKEY	t3=[rk1],2*KSZ		// 1/3:rk[3]
+	add	te30=te30,te0		// 1/1:te0+s0
+	shr.u	te01=s1,twenty4	};;	// 1/1:s1>>24
+{ .mmi;	ld1	te33=[te33]		// 2/0:te0[s3&0xff]
+	add	te22=te22,te0		// 2/0:te0+s2>>8&0xff
+	extr.u	te20=s0,8,8	}	// 2/2:s0>>8&0xff
+{ .mmi;	ld1	te30=[te30]		// 2/1:te0[s0]
+	add	te23=te23,te0		// 2/1:te0+s3>>8
+	shr.u	te02=s2,twenty4	};;	// 2/2:s2>>24
+{ .mmi;	ld1	te22=[te22]		// 3/0:te0[s2>>8]
+	add	te20=te20,te0		// 3/2:te0+s0>>8
+	extr.u	te21=s1,8,8	}	// 3/3:s1>>8&0xff
+{ .mmi;	ld1	te23=[te23]		// 3/1:te0[s3>>8]
+	add	te00=te00,te0		// 3/0:te0+s0>>24
+	shr.u	te03=s3,twenty4	};;	// 3/3:s3>>24
+{ .mmi;	ld1	te20=[te20]		// 4/2:te0[s0>>8]
+	add	te21=te21,te0		// 4/3:te0+s2
+	extr.u	te11=s1,16,8	}	// 4/0:s1>>16&0xff
+{ .mmi;	ld1	te00=[te00]		// 4/0:te0[s0>>24]
+	add	te01=te01,te0		// 4/1:te0+s1>>24
+	shr.u	te13=s3,sixteen	};;	// 4/2:s3>>16
+{ .mmi;	ld1	te21=[te21]		// 5/3:te0[s1>>8]
+	add	te11=te11,te0		// 5/0:te0+s1>>16
+	extr.u	te12=s2,16,8	}	// 5/1:s2>>16&0xff
+{ .mmi;	ld1	te01=[te01]		// 5/1:te0[s1>>24]
+	add	te02=te02,te0		// 5/2:te0+s2>>24
+	and	te31=s1,maskff	};;	// 5/2:s1&0xff
+{ .mmi;	ld1	te11=[te11]		// 6/0:te0[s1>>16]
+	add	te12=te12,te0		// 6/1:te0+s2>>16
+	extr.u	te10=s0,16,8	}	// 6/3:s0>>16&0xff
+{ .mmi;	ld1	te02=[te02]		// 6/2:te0[s2>>24]
+	add	te03=te03,te0		// 6/3:te0+s0>>16
+	and	te32=s2,maskff	};;	// 6/3:s2&0xff
+
+{ .mmi;	ld1	te12=[te12]		// 7/1:te0[s2>>16]
+	add	te31=te31,te0		// 7/2:te0+s1&0xff
+	dep	te33=te22,te33,8,8}	// 7/0:
+{ .mmi;	ld1	te03=[te03]		// 7/3:te0[s3>>24]
+	add	te32=te32,te0		// 7/3:te0+s2
+	and	te13=te13,maskff};;	// 7/2:s3>>16&0xff
+{ .mmi;	ld1	te31=[te31]		// 8/2:te0[s1]
+	add	te13=te13,te0		// 8/2:te0+s3>>16
+	dep	te30=te23,te30,8,8}	// 8/1:
+{ .mmi;	ld1	te32=[te32]		// 8/3:te0[s2]
+	add	te10=te10,te0		// 8/3:te0+s0>>16
+	shl	te00=te00,twenty4};;	// 8/0:
+{ .mii;	ld1	te13=[te13]		// 9/2:te0[s3>>16]
+	dep	te33=te11,te33,16,8	// 9/0:
+	shl	te01=te01,twenty4};;	// 9/1:
+{ .mii;	ld1	te10=[te10]		// 10/3:te0[s0>>16]
+	dep	te31=te20,te31,8,8	// 10/2:
+	shl	te02=te02,twenty4};;	// 10/2:
+{ .mii;	xor	t0=t0,te33		// 11/0:
+	dep	te32=te21,te32,8,8	// 11/3:
+	shl	te12=te12,sixteen};;	// 11/1:
+{ .mii;	xor	r16=t0,te00		// 12/0:done!
+	dep	te31=te13,te31,16,8	// 12/2:
+	shl	te03=te03,twenty4};;	// 12/3:
+{ .mmi;	xor	t1=t1,te01		// 13/1:
+	xor	t2=t2,te02		// 13/2:
+	dep	te32=te10,te32,16,8};;	// 13/3:
+{ .mmi;	xor	t1=t1,te30		// 14/1:
+	xor	r24=t2,te31		// 14/2:done!
+	xor	t3=t3,te32	};;	// 14/3:
+{ .mib;	xor	r20=t1,te12		// 15/1:done!
+	xor	r28=t3,te03		// 15/3:done!
+	br.ret.sptk	b6	};;
 .endp	_ia64_AES_encrypt#
 
 // void AES_encrypt (const void *in,void *out,const AES_KEY *key);
 .global	AES_encrypt#
 .proc	AES_encrypt#
 .align	32
-.skip	16
 AES_encrypt:
 	.prologue
-	.fframe	0
-	.save	ar.pfs,r2
-	.save	ar.lc,r3
-{ .mmi;	alloc	r2=ar.pfs,3,0,12,0
-	addl	out8=@ltoff(AES_Te#),gp
-	mov	r3=ar.lc		}
-{ .mmi;	and	out0=3,in0
-	ADDP	in0=0,in0
+	.save	ar.pfs,pfssave
+{ .mmi;	alloc	pfssave=ar.pfs,3,1,12,0
+	and	out0=3,in0
+	mov	r3=ip			}
+{ .mmi;	ADDP	in0=0,in0
+	mov	loc0=psr.um
 	ADDP	out11=KSZ*60,in2	};;	// &AES_KEY->rounds
 
-	.body
-{ .mmi;	ld8	out8=[out8]			// Te0
-	ld4	out11=[out11]			// AES_KEY->rounds
+{ .mmi;	ld4	out11=[out11]			// AES_KEY->rounds
+	add	out8=(AES_Te#-AES_encrypt#),r3	// Te0
+	.save	pr,prsave
 	mov	prsave=pr		}
+{ .mmi;	rum	1<<3				// clear um.ac
+	.save	ar.lc,lcsave
+	mov	lcsave=ar.lc		};;
 
+	.body
 #if defined(_HPUX_SOURCE)	// HPUX is big-endian, cut 15+15 cycles...
 { .mib; cmp.ne	p6,p0=out0,r0
 	add	out0=4,in0
@@ -216,8 +317,9 @@ AES_encrypt:
 	ADDP	in1=0,in1
 (p6)	br.spnt	.Le_o_unaligned		};;
 
-{ .mii;	mov	ar.pfs=r2
-	mov	ar.lc=r3		}
+{ .mii;	mov	psr.um=loc0
+	mov	ar.pfs=pfssave
+	mov	ar.lc=lcsave		};;
 { .mmi;	st4	[in1]=r16,8		// s0
 	st4	[in0]=r20,8		// s1
 	mov	pr=prsave,0x1ffff	};;
@@ -278,13 +380,13 @@ AES_encrypt:
 	shr.u	r23=r20,twenty4		}//;;	// s1
 { .mii;	ADDP	out2=2,in1
 	extr.u	r21=r20,8,8
-	shr.u	r22=r20,sixteen	}//;;
+	shr.u	r22=r20,sixteen		}//;;
 { .mii;	ADDP	out3=3,in1
 	extr.u	r25=r24,8,8			// s2
 	shr.u	r27=r24,twenty4		};;
 { .mii;	st1	[out3]=r16,4
 	extr.u	r26=r24,16,8
-	shr.u	r31=r28,twenty4	}//;;	// s3
+	shr.u	r31=r28,twenty4		}//;;	// s3
 { .mii;	st1	[out2]=r17,4
 	extr.u	r29=r28,8,8
 	shr.u	r30=r28,sixteen		}//;;
@@ -300,12 +402,13 @@ AES_encrypt:
 	mov	pr=prsave,0x1ffff	}//;;
 { .mmi;	st1	[out1]=r26,4
 	st1	[out0]=r27,4
-	mov	ar.pfs=r2		};;
+	mov	ar.pfs=pfssave		};;
 { .mmi;	st1	[out3]=r28
 	st1	[out2]=r29
-	mov	ar.lc=r3		}//;;
-{ .mmb;	st1	[out1]=r30
-	st1	[out0]=r31
+	mov	ar.lc=lcsave		}//;;
+{ .mmi;	st1	[out1]=r30
+	st1	[out0]=r31		}
+{ .mfb;	mov	psr.um=loc0			// restore user mask
 	br.ret.sptk.many	b0	};;
 .endp	AES_encrypt#
 
@@ -360,16 +463,19 @@ while(<>) {
 // Clobber:	r16-r31,rk0-rk1,r32-r43
 .align	32
 _ia64_AES_decrypt:
+	.prologue
+	.altrp	b6
+	.body
 { .mmi;	alloc	r16=ar.pfs,12,0,0,8
 	LDKEY	t0=[rk0],2*KSZ
 	mov	pr.rot=1<<16	}
 { .mmi;	LDKEY	t1=[rk1],2*KSZ
-	add	te1=1024,te0
+	add	te1=TE1,te0
 	add	te3=-3,te3	};;
 { .mib;	LDKEY	t2=[rk0],2*KSZ
-	mov	ar.ec=3		}
+	mov	ar.ec=2		}
 { .mib;	LDKEY	t3=[rk1],2*KSZ
-	add	te2=2048,te0
+	add	te2=TE2,te0
 	brp.loop.imp	.Ld_top,.Ld_end-16	};;
 
 { .mmi;	xor	s0=s0,t0
@@ -377,8 +483,8 @@ _ia64_AES_decrypt:
 	mov	ar.lc=te3	}
 { .mmi;	xor	s2=s2,t2
 	xor	s3=s3,t3
-	add	te3=3072,te0	};;
-
+	add	te3=TE3,te0	};;
+
 .align	32
 .Ld_top:
 { .mmi;	(p0)	LDKEY	t0=[rk0],2*KSZ		// 0/0:rk[0]
@@ -388,105 +494,184 @@ _ia64_AES_decrypt:
 	(p0)	and	te32=s2,maskff		// 0/1:s0&0xff
 	(p0)	shr.u	te00=s0,twenty4	};;	// 0/0:s0>>24
 { .mmi;	(p0)	LDKEY	t2=[rk0],2*KSZ		// 1/2:rk[2]
-	(p0)	shladd	te31=te31,2,te3		// 1/0:te0+s0>>24
+	(p0)	shladd	te31=te31,3,te3		// 1/0:te0+s0>>24
 	(p0)	extr.u	te23=s3,8,8	}	// 1/1:s3>>8&0xff
 { .mmi;	(p0)	LDKEY	t3=[rk1],2*KSZ		// 1/3:rk[3]
-	(p0)	shladd	te32=te32,2,te3		// 1/1:te3+s0
+	(p0)	shladd	te32=te32,3,te3		// 1/1:te3+s0
 	(p0)	shr.u	te01=s1,twenty4	};;	// 1/1:s1>>24
 { .mmi;	(p0)	ld4	te31=[te31]		// 2/0:te3[s3&0xff]
-	(p0)	shladd	te22=te22,2,te2		// 2/0:te2+s2>>8&0xff
+	(p0)	shladd	te22=te22,3,te2		// 2/0:te2+s2>>8&0xff
 	(p0)	extr.u	te20=s0,8,8	}	// 2/2:s0>>8&0xff
 { .mmi;	(p0)	ld4	te32=[te32]		// 2/1:te3[s0]
-	(p0)	shladd	te23=te23,2,te2		// 2/1:te2+s3>>8
+	(p0)	shladd	te23=te23,3,te2		// 2/1:te2+s3>>8
 	(p0)	shr.u	te02=s2,twenty4	};;	// 2/2:s2>>24
 { .mmi;	(p0)	ld4	te22=[te22]		// 3/0:te2[s2>>8]
-	(p0)	shladd	te20=te20,2,te2		// 3/2:te2+s0>>8
+	(p0)	shladd	te20=te20,3,te2		// 3/2:te2+s0>>8
 	(p0)	extr.u	te21=s1,8,8	}	// 3/3:s1>>8&0xff
 { .mmi;	(p0)	ld4	te23=[te23]		// 3/1:te2[s3>>8]
-	(p0)	shladd	te00=te00,2,te0		// 3/0:te0+s0>>24
+	(p0)	shladd	te00=te00,3,te0		// 3/0:te0+s0>>24
 	(p0)	shr.u	te03=s3,twenty4	};;	// 3/3:s3>>24
 { .mmi;	(p0)	ld4	te20=[te20]		// 4/2:te2[s0>>8]
-	(p0)	shladd	te21=te21,2,te2		// 4/3:te3+s2
+	(p0)	shladd	te21=te21,3,te2		// 4/3:te3+s2
 	(p0)	extr.u	te13=s3,16,8	}	// 4/0:s1>>16&0xff
 { .mmi;	(p0)	ld4	te00=[te00]		// 4/0:te0[s0>>24]
-	(p0)	shladd	te01=te01,2,te0		// 4/1:te0+s1>>24
+	(p0)	shladd	te01=te01,3,te0		// 4/1:te0+s1>>24
 	(p0)	shr.u	te11=s1,sixteen	};;	// 4/2:s3>>16
 { .mmi;	(p0)	ld4	te21=[te21]		// 5/3:te2[s1>>8]
-	(p0)	shladd	te13=te13,2,te1		// 5/0:te1+s1>>16
+	(p0)	shladd	te13=te13,3,te1		// 5/0:te1+s1>>16
 	(p0)	extr.u	te10=s0,16,8	}	// 5/1:s2>>16&0xff
 { .mmi;	(p0)	ld4	te01=[te01]		// 5/1:te0[s1>>24]
-	(p0)	shladd	te02=te02,2,te0		// 5/2:te0+s2>>24
+	(p0)	shladd	te02=te02,3,te0		// 5/2:te0+s2>>24
 	(p0)	and	te33=s3,maskff	};;	// 5/2:s1&0xff
-
 { .mmi;	(p0)	ld4	te13=[te13]		// 6/0:te1[s1>>16]
-	(p0)	shladd	te10=te10,2,te1		// 6/1:te1+s2>>16
+	(p0)	shladd	te10=te10,3,te1		// 6/1:te1+s2>>16
 	(p0)	extr.u	te12=s2,16,8	}	// 6/3:s0>>16&0xff
 { .mmi;	(p0)	ld4	te02=[te02]		// 6/2:te0[s2>>24]
-	(p0)	shladd	te03=te03,2,te0		// 6/3:te1+s0>>16
+	(p0)	shladd	te03=te03,3,te0		// 6/3:te1+s0>>16
 	(p0)	and	te30=s0,maskff	};;	// 6/3:s2&0xff
+
 { .mmi;	(p0)	ld4	te10=[te10]		// 7/1:te1[s2>>16]
-	(p0)	shladd	te33=te33,2,te3		// 7/2:te3+s1&0xff
+	(p0)	shladd	te33=te33,3,te3		// 7/2:te3+s1&0xff
 	(p0)	and	te11=te11,maskff}	// 7/2:s3>>16&0xff
 { .mmi;	(p0)	ld4	te03=[te03]		// 7/3:te0[s3>>24]
-	(p0)	shladd	te30=te30,2,te3		// 7/3:te3+s2
+	(p0)	shladd	te30=te30,3,te3		// 7/3:te3+s2
 	(p0)	xor	t0=t0,te31	};;	// 7/0:
 { .mmi;	(p0)	ld4	te33=[te33]		// 8/2:te3[s1]
-	(p0)	shladd	te11=te11,2,te1		// 8/2:te1+s3>>16
+	(p0)	shladd	te11=te11,3,te1		// 8/2:te1+s3>>16
 	(p0)	xor	t0=t0,te22	}	// 8/0:
 { .mmi;	(p0)	ld4	te30=[te30]		// 8/3:te3[s2]
-	(p0)	shladd	te12=te12,2,te1		// 8/3:te1+s0>>16
+	(p0)	shladd	te12=te12,3,te1		// 8/3:te1+s0>>16
 	(p0)	xor	t1=t1,te32	};;	// 8/1:
 { .mmi;	(p0)	ld4	te11=[te11]		// 9/2:te1[s3>>16]
-	(p0)	xor	t0=t0,te00		// 9/0:
-	(p0)	xor	t1=t1,te23	}	// 9/1:		
-{ .mmi;	(p0)	ld4	te12=[te12]		// 9/3:te1[s0>>16]
-	(p0)	xor	t2=t2,te20		// 9/2:
-	(p0)	xor	t3=t3,te21	};;	// 9/3:
-{ .mmi;	(p0)	xor	t0=t0,te13		// 10/0:done!
-	(p0)	xor	t1=t1,te01		// 10/1:
-	(p0)	xor	t2=t2,te02	}	// 10/2:
-{ .mmi;	(p0)	xor	t3=t3,te03		// 10/3:
-	(p16)	cmp.eq	p0,p17=r0,r0 	};;	// 10/clear (p17)
-{ .mmi;	(p0)	xor	t1=t1,te10		// 11/1:done!
-	(p0)	xor	t2=t2,te33		// 11/2:
-	(p0)	xor	t3=t3,te30	}	// 11/3:
-{ .mmi;	(p17)	add	te0=4096,te0		// 11/	
-	(p17)	add	te1=4096,te1	};;	// 11/
-{ .mib;	(p0)	xor	t2=t2,te11		// 12/2:done!
-	(p0)	xor	t3=t3,te12	}	// 12/3:done!
-{ .mib;	(p17)	add	te2=4096,te2		// 12/
-	(p17)	add	te3=4096,te3		// 12/
+	(p0)	ld4	te12=[te12]		// 9/3:te1[s0>>16]
+	(p0)	xor	t0=t0,te00	};;	// 9/0:		!L2 scheduling
+{ .mmi;	(p0)	xor	t1=t1,te23		// 10[9]/1:	
+	(p0)	xor	t2=t2,te20		// 10[9]/2:
+	(p0)	xor	t3=t3,te21	};;	// 10[9]/3:
+{ .mmi;	(p0)	xor	t0=t0,te13		// 11[10]/0:done!
+	(p0)	xor	t1=t1,te01		// 11[10]/1:
+	(p0)	xor	t2=t2,te02	};;	// 11[10]/2:	!L2 scheduling
+{ .mmi;	(p0)	xor	t3=t3,te03		// 12[10]/3:
+	(p16)	cmp.eq	p0,p17=r0,r0 	};;	// 12[10]/clear (p17)
+{ .mmi;	(p0)	xor	t1=t1,te10		// 13[11]/1:done!
+	(p0)	xor	t2=t2,te33		// 13[11]/2:
+	(p0)	xor	t3=t3,te30	}	// 13[11]/3:
+{ .mmi;	(p17)	add	te0=2048,te0		// 13[11]/
+	(p17)	add	te1=2048+64-TE1,te1};;	// 13[11]/
+{ .mib;	(p0)	xor	t2=t2,te11		// 14[12]/2:done!
+	(p17)	add	te2=2048+128-TE2,te2}	// 14[12]/
+{ .mib;	(p0)	xor	t3=t3,te12		// 14[12]/3:done!
+	(p17)	add	te3=2048+192-TE3,te3	// 14[12]/
 	br.ctop.sptk	.Ld_top		};;
 .Ld_end:
-{ .mib;	mov	r16=s0
-	mov	r20=s1			}
-{ .mib;	mov	r24=s2
-	mov	r28=s3
-	br.ret.sptk	b6		};;
+
+
+{ .mmi;	ld8	te10=[te0]		// prefetch Td4
+	ld8	te33=[te1]	}
+{ .mmi;	ld8	te12=[te2]
+	ld8	te30=[te3]	}
+
+{ .mmi;	LDKEY	t0=[rk0],2*KSZ		// 0/0:rk[0]
+	and	te31=s1,maskff		// 0/0:s3&0xff
+	extr.u	te22=s2,8,8	}	// 0/0:s2>>8&0xff
+{ .mmi; LDKEY	t1=[rk1],2*KSZ		// 0/1:rk[1]
+	and	te32=s2,maskff		// 0/1:s0&0xff
+	shr.u	te00=s0,twenty4	};;	// 0/0:s0>>24
+{ .mmi;	LDKEY	t2=[rk0],2*KSZ		// 1/2:rk[2]
+	add	te31=te31,te0		// 1/0:te0+s0>>24
+	extr.u	te23=s3,8,8	}	// 1/1:s3>>8&0xff
+{ .mmi;	LDKEY	t3=[rk1],2*KSZ		// 1/3:rk[3]
+	add	te32=te32,te0		// 1/1:te0+s0
+	shr.u	te01=s1,twenty4	};;	// 1/1:s1>>24
+{ .mmi;	ld1	te31=[te31]		// 2/0:te0[s3&0xff]
+	add	te22=te22,te0		// 2/0:te0+s2>>8&0xff
+	extr.u	te20=s0,8,8	}	// 2/2:s0>>8&0xff
+{ .mmi;	ld1	te32=[te32]		// 2/1:te0[s0]
+	add	te23=te23,te0		// 2/1:te0+s3>>8
+	shr.u	te02=s2,twenty4	};;	// 2/2:s2>>24
+{ .mmi;	ld1	te22=[te22]		// 3/0:te0[s2>>8]
+	add	te20=te20,te0		// 3/2:te0+s0>>8
+	extr.u	te21=s1,8,8	}	// 3/3:s1>>8&0xff
+{ .mmi;	ld1	te23=[te23]		// 3/1:te0[s3>>8]
+	add	te00=te00,te0		// 3/0:te0+s0>>24
+	shr.u	te03=s3,twenty4	};;	// 3/3:s3>>24
+{ .mmi;	ld1	te20=[te20]		// 4/2:te0[s0>>8]
+	add	te21=te21,te0		// 4/3:te0+s2
+	extr.u	te13=s3,16,8	}	// 4/0:s1>>16&0xff
+{ .mmi;	ld1	te00=[te00]		// 4/0:te0[s0>>24]
+	add	te01=te01,te0		// 4/1:te0+s1>>24
+	shr.u	te11=s1,sixteen	};;	// 4/2:s3>>16
+{ .mmi;	ld1	te21=[te21]		// 5/3:te0[s1>>8]
+	add	te13=te13,te0		// 5/0:te0+s1>>16
+	extr.u	te10=s0,16,8	}	// 5/1:s2>>16&0xff
+{ .mmi;	ld1	te01=[te01]		// 5/1:te0[s1>>24]
+	add	te02=te02,te0		// 5/2:te0+s2>>24
+	and	te33=s3,maskff	};;	// 5/2:s1&0xff
+{ .mmi;	ld1	te13=[te13]		// 6/0:te0[s1>>16]
+	add	te10=te10,te0		// 6/1:te0+s2>>16
+	extr.u	te12=s2,16,8	}	// 6/3:s0>>16&0xff
+{ .mmi;	ld1	te02=[te02]		// 6/2:te0[s2>>24]
+	add	te03=te03,te0		// 6/3:te0+s0>>16
+	and	te30=s0,maskff	};;	// 6/3:s2&0xff
+
+{ .mmi;	ld1	te10=[te10]		// 7/1:te0[s2>>16]
+	add	te33=te33,te0		// 7/2:te0+s1&0xff
+	dep	te31=te22,te31,8,8}	// 7/0:
+{ .mmi;	ld1	te03=[te03]		// 7/3:te0[s3>>24]
+	add	te30=te30,te0		// 7/3:te0+s2
+	and	te11=te11,maskff};;	// 7/2:s3>>16&0xff
+{ .mmi;	ld1	te33=[te33]		// 8/2:te0[s1]
+	add	te11=te11,te0		// 8/2:te0+s3>>16
+	dep	te32=te23,te32,8,8}	// 8/1:
+{ .mmi;	ld1	te30=[te30]		// 8/3:te0[s2]
+	add	te12=te12,te0		// 8/3:te0+s0>>16
+	shl	te00=te00,twenty4};;	// 8/0:
+{ .mii;	ld1	te11=[te11]		// 9/2:te0[s3>>16]
+	dep	te31=te13,te31,16,8	// 9/0:
+	shl	te01=te01,twenty4};;	// 9/1:
+{ .mii;	ld1	te12=[te12]		// 10/3:te0[s0>>16]
+	dep	te33=te20,te33,8,8	// 10/2:
+	shl	te02=te02,twenty4};;	// 10/2:
+{ .mii;	xor	t0=t0,te31		// 11/0:
+	dep	te30=te21,te30,8,8	// 11/3:
+	shl	te10=te10,sixteen};;	// 11/1:
+{ .mii;	xor	r16=t0,te00		// 12/0:done!
+	dep	te33=te11,te33,16,8	// 12/2:
+	shl	te03=te03,twenty4};;	// 12/3:
+{ .mmi;	xor	t1=t1,te01		// 13/1:
+	xor	t2=t2,te02		// 13/2:
+	dep	te30=te12,te30,16,8};;	// 13/3:
+{ .mmi;	xor	t1=t1,te32		// 14/1:
+	xor	r24=t2,te33		// 14/2:done!
+	xor	t3=t3,te30	};;	// 14/3:
+{ .mib;	xor	r20=t1,te10		// 15/1:done!
+	xor	r28=t3,te03		// 15/3:done!
+	br.ret.sptk	b6	};;
 .endp	_ia64_AES_decrypt#
 
 // void AES_decrypt (const void *in,void *out,const AES_KEY *key);
 .global	AES_decrypt#
 .proc	AES_decrypt#
 .align	32
-.skip	16
 AES_decrypt:
 	.prologue
-	.fframe	0
-	.save	ar.pfs,r2
-	.save	ar.lc,r3
-{ .mmi;	alloc	r2=ar.pfs,3,0,12,0
-	addl	out8=@ltoff(AES_Td#),gp
-	mov	r3=ar.lc		}
-{ .mmi;	and	out0=3,in0
-	ADDP	in0=0,in0
+	.save	ar.pfs,pfssave
+{ .mmi;	alloc	pfssave=ar.pfs,3,1,12,0
+	and	out0=3,in0
+	mov	r3=ip			}
+{ .mmi;	ADDP	in0=0,in0
+	mov	loc0=psr.um
 	ADDP	out11=KSZ*60,in2	};;	// &AES_KEY->rounds
 
-	.body
-{ .mmi;	ld8	out8=[out8]			// Te0
-	ld4	out11=[out11]			// AES_KEY->rounds
+{ .mmi;	ld4	out11=[out11]			// AES_KEY->rounds
+	add	out8=(AES_Td#-AES_decrypt#),r3	// Te0
+	.save	pr,prsave
 	mov	prsave=pr		}
+{ .mmi;	rum	1<<3				// clear um.ac
+	.save	ar.lc,lcsave
+	mov	lcsave=ar.lc		};;
 
+	.body
 #if defined(_HPUX_SOURCE)	// HPUX is big-endian, cut 15+15 cycles...
 { .mib; cmp.ne	p6,p0=out0,r0
 	add	out0=4,in0
@@ -509,8 +694,9 @@ AES_decrypt:
 	ADDP	in1=0,in1
 (p6)	br.spnt	.Ld_o_unaligned		};;
 
-{ .mii;	mov	ar.pfs=r2
-	mov	ar.lc=r3		}
+{ .mii;	mov	psr.um=loc0
+	mov	ar.pfs=pfssave
+	mov	ar.lc=lcsave		};;
 { .mmi;	st4	[in1]=r16,8		// s0
 	st4	[in0]=r20,8		// s1
 	mov	pr=prsave,0x1ffff	};;
@@ -571,13 +757,13 @@ AES_decrypt:
 	shr.u	r23=r20,twenty4		}//;;	// s1
 { .mii;	ADDP	out2=2,in1
 	extr.u	r21=r20,8,8
-	shr.u	r22=r20,sixteen	}//;;
+	shr.u	r22=r20,sixteen		}//;;
 { .mii;	ADDP	out3=3,in1
 	extr.u	r25=r24,8,8			// s2
 	shr.u	r27=r24,twenty4		};;
 { .mii;	st1	[out3]=r16,4
 	extr.u	r26=r24,16,8
-	shr.u	r31=r28,twenty4	}//;;	// s3
+	shr.u	r31=r28,twenty4		}//;;	// s3
 { .mii;	st1	[out2]=r17,4
 	extr.u	r29=r28,8,8
 	shr.u	r30=r28,sixteen		}//;;
@@ -593,12 +779,13 @@ AES_decrypt:
 	mov	pr=prsave,0x1ffff	}//;;
 { .mmi;	st1	[out1]=r26,4
 	st1	[out0]=r27,4
-	mov	ar.pfs=r2		};;
+	mov	ar.pfs=pfssave		};;
 { .mmi;	st1	[out3]=r28
 	st1	[out2]=r29
-	mov	ar.lc=r3		}//;;
-{ .mmb;	st1	[out1]=r30
-	st1	[out0]=r31
+	mov	ar.lc=lcsave		}//;;
+{ .mmi;	st1	[out1]=r30
+	st1	[out0]=r31		}
+{ .mfb;	mov	psr.um=loc0			// restore user mask
 	br.ret.sptk.many	b0	};;
 .endp	AES_decrypt#
 
@@ -606,1047 +793,331 @@ AES_decrypt:
 .align	64
 .global	AES_Te#
 .type	AES_Te#,@object
-AES_Te:	data4	0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d
-	data4	0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554
-	data4	0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d
-	data4	0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a
-	data4	0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87
-	data4	0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b
-	data4	0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea
-	data4	0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b
-	data4	0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a
-	data4	0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f
-	data4	0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108
-	data4	0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f
-	data4	0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e
-	data4	0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5
-	data4	0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d
-	data4	0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f
-	data4	0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e
-	data4	0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb
-	data4	0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce
-	data4	0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497
-	data4	0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c
-	data4	0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed
-	data4	0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b
-	data4	0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a
-	data4	0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16
-	data4	0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594
-	data4	0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81
-	data4	0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3
-	data4	0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a
-	data4	0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504
-	data4	0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163
-	data4	0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d
-	data4	0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f
-	data4	0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739
-	data4	0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47
-	data4	0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395
-	data4	0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f
-	data4	0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883
-	data4	0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c
-	data4	0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76
-	data4	0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e
-	data4	0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4
-	data4	0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6
-	data4	0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b
-	data4	0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7
-	data4	0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0
-	data4	0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25
-	data4	0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818
-	data4	0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72
-	data4	0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651
-	data4	0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21
-	data4	0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85
-	data4	0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa
-	data4	0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12
-	data4	0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0
-	data4	0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9
-	data4	0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133
-	data4	0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7
-	data4	0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920
-	data4	0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a
-	data4	0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17
-	data4	0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8
-	data4	0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11
-	data4	0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a
-// Te1:
-	data4	0xa5c66363, 0x84f87c7c, 0x99ee7777, 0x8df67b7b
-	data4	0x0dfff2f2, 0xbdd66b6b, 0xb1de6f6f, 0x5491c5c5
-	data4	0x50603030, 0x03020101, 0xa9ce6767, 0x7d562b2b
-	data4	0x19e7fefe, 0x62b5d7d7, 0xe64dabab, 0x9aec7676
-	data4	0x458fcaca, 0x9d1f8282, 0x4089c9c9, 0x87fa7d7d
-	data4	0x15effafa, 0xebb25959, 0xc98e4747, 0x0bfbf0f0
-	data4	0xec41adad, 0x67b3d4d4, 0xfd5fa2a2, 0xea45afaf
-	data4	0xbf239c9c, 0xf753a4a4, 0x96e47272, 0x5b9bc0c0
-	data4	0xc275b7b7, 0x1ce1fdfd, 0xae3d9393, 0x6a4c2626
-	data4	0x5a6c3636, 0x417e3f3f, 0x02f5f7f7, 0x4f83cccc
-	data4	0x5c683434, 0xf451a5a5, 0x34d1e5e5, 0x08f9f1f1
-	data4	0x93e27171, 0x73abd8d8, 0x53623131, 0x3f2a1515
-	data4	0x0c080404, 0x5295c7c7, 0x65462323, 0x5e9dc3c3
-	data4	0x28301818, 0xa1379696, 0x0f0a0505, 0xb52f9a9a
-	data4	0x090e0707, 0x36241212, 0x9b1b8080, 0x3ddfe2e2
-	data4	0x26cdebeb, 0x694e2727, 0xcd7fb2b2, 0x9fea7575
-	data4	0x1b120909, 0x9e1d8383, 0x74582c2c, 0x2e341a1a
-	data4	0x2d361b1b, 0xb2dc6e6e, 0xeeb45a5a, 0xfb5ba0a0
-	data4	0xf6a45252, 0x4d763b3b, 0x61b7d6d6, 0xce7db3b3
-	data4	0x7b522929, 0x3edde3e3, 0x715e2f2f, 0x97138484
-	data4	0xf5a65353, 0x68b9d1d1, 0x00000000, 0x2cc1eded
-	data4	0x60402020, 0x1fe3fcfc, 0xc879b1b1, 0xedb65b5b
-	data4	0xbed46a6a, 0x468dcbcb, 0xd967bebe, 0x4b723939
-	data4	0xde944a4a, 0xd4984c4c, 0xe8b05858, 0x4a85cfcf
-	data4	0x6bbbd0d0, 0x2ac5efef, 0xe54faaaa, 0x16edfbfb
-	data4	0xc5864343, 0xd79a4d4d, 0x55663333, 0x94118585
-	data4	0xcf8a4545, 0x10e9f9f9, 0x06040202, 0x81fe7f7f
-	data4	0xf0a05050, 0x44783c3c, 0xba259f9f, 0xe34ba8a8
-	data4	0xf3a25151, 0xfe5da3a3, 0xc0804040, 0x8a058f8f
-	data4	0xad3f9292, 0xbc219d9d, 0x48703838, 0x04f1f5f5
-	data4	0xdf63bcbc, 0xc177b6b6, 0x75afdada, 0x63422121
-	data4	0x30201010, 0x1ae5ffff, 0x0efdf3f3, 0x6dbfd2d2
-	data4	0x4c81cdcd, 0x14180c0c, 0x35261313, 0x2fc3ecec
-	data4	0xe1be5f5f, 0xa2359797, 0xcc884444, 0x392e1717
-	data4	0x5793c4c4, 0xf255a7a7, 0x82fc7e7e, 0x477a3d3d
-	data4	0xacc86464, 0xe7ba5d5d, 0x2b321919, 0x95e67373
-	data4	0xa0c06060, 0x98198181, 0xd19e4f4f, 0x7fa3dcdc
-	data4	0x66442222, 0x7e542a2a, 0xab3b9090, 0x830b8888
-	data4	0xca8c4646, 0x29c7eeee, 0xd36bb8b8, 0x3c281414
-	data4	0x79a7dede, 0xe2bc5e5e, 0x1d160b0b, 0x76addbdb
-	data4	0x3bdbe0e0, 0x56643232, 0x4e743a3a, 0x1e140a0a
-	data4	0xdb924949, 0x0a0c0606, 0x6c482424, 0xe4b85c5c
-	data4	0x5d9fc2c2, 0x6ebdd3d3, 0xef43acac, 0xa6c46262
-	data4	0xa8399191, 0xa4319595, 0x37d3e4e4, 0x8bf27979
-	data4	0x32d5e7e7, 0x438bc8c8, 0x596e3737, 0xb7da6d6d
-	data4	0x8c018d8d, 0x64b1d5d5, 0xd29c4e4e, 0xe049a9a9
-	data4	0xb4d86c6c, 0xfaac5656, 0x07f3f4f4, 0x25cfeaea
-	data4	0xafca6565, 0x8ef47a7a, 0xe947aeae, 0x18100808
-	data4	0xd56fbaba, 0x88f07878, 0x6f4a2525, 0x725c2e2e
-	data4	0x24381c1c, 0xf157a6a6, 0xc773b4b4, 0x5197c6c6
-	data4	0x23cbe8e8, 0x7ca1dddd, 0x9ce87474, 0x213e1f1f
-	data4	0xdd964b4b, 0xdc61bdbd, 0x860d8b8b, 0x850f8a8a
-	data4	0x90e07070, 0x427c3e3e, 0xc471b5b5, 0xaacc6666
-	data4	0xd8904848, 0x05060303, 0x01f7f6f6, 0x121c0e0e
-	data4	0xa3c26161, 0x5f6a3535, 0xf9ae5757, 0xd069b9b9
-	data4	0x91178686, 0x5899c1c1, 0x273a1d1d, 0xb9279e9e
-	data4	0x38d9e1e1, 0x13ebf8f8, 0xb32b9898, 0x33221111
-	data4	0xbbd26969, 0x70a9d9d9, 0x89078e8e, 0xa7339494
-	data4	0xb62d9b9b, 0x223c1e1e, 0x92158787, 0x20c9e9e9
-	data4	0x4987cece, 0xffaa5555, 0x78502828, 0x7aa5dfdf
-	data4	0x8f038c8c, 0xf859a1a1, 0x80098989, 0x171a0d0d
-	data4	0xda65bfbf, 0x31d7e6e6, 0xc6844242, 0xb8d06868
-	data4	0xc3824141, 0xb0299999, 0x775a2d2d, 0x111e0f0f
-	data4	0xcb7bb0b0, 0xfca85454, 0xd66dbbbb, 0x3a2c1616
-// Te2:
-	data4	0x63a5c663, 0x7c84f87c, 0x7799ee77, 0x7b8df67b
-	data4	0xf20dfff2, 0x6bbdd66b, 0x6fb1de6f, 0xc55491c5
-	data4	0x30506030, 0x01030201, 0x67a9ce67, 0x2b7d562b
-	data4	0xfe19e7fe, 0xd762b5d7, 0xabe64dab, 0x769aec76
-	data4	0xca458fca, 0x829d1f82, 0xc94089c9, 0x7d87fa7d
-	data4	0xfa15effa, 0x59ebb259, 0x47c98e47, 0xf00bfbf0
-	data4	0xadec41ad, 0xd467b3d4, 0xa2fd5fa2, 0xafea45af
-	data4	0x9cbf239c, 0xa4f753a4, 0x7296e472, 0xc05b9bc0
-	data4	0xb7c275b7, 0xfd1ce1fd, 0x93ae3d93, 0x266a4c26
-	data4	0x365a6c36, 0x3f417e3f, 0xf702f5f7, 0xcc4f83cc
-	data4	0x345c6834, 0xa5f451a5, 0xe534d1e5, 0xf108f9f1
-	data4	0x7193e271, 0xd873abd8, 0x31536231, 0x153f2a15
-	data4	0x040c0804, 0xc75295c7, 0x23654623, 0xc35e9dc3
-	data4	0x18283018, 0x96a13796, 0x050f0a05, 0x9ab52f9a
-	data4	0x07090e07, 0x12362412, 0x809b1b80, 0xe23ddfe2
-	data4	0xeb26cdeb, 0x27694e27, 0xb2cd7fb2, 0x759fea75
-	data4	0x091b1209, 0x839e1d83, 0x2c74582c, 0x1a2e341a
-	data4	0x1b2d361b, 0x6eb2dc6e, 0x5aeeb45a, 0xa0fb5ba0
-	data4	0x52f6a452, 0x3b4d763b, 0xd661b7d6, 0xb3ce7db3
-	data4	0x297b5229, 0xe33edde3, 0x2f715e2f, 0x84971384
-	data4	0x53f5a653, 0xd168b9d1, 0x00000000, 0xed2cc1ed
-	data4	0x20604020, 0xfc1fe3fc, 0xb1c879b1, 0x5bedb65b
-	data4	0x6abed46a, 0xcb468dcb, 0xbed967be, 0x394b7239
-	data4	0x4ade944a, 0x4cd4984c, 0x58e8b058, 0xcf4a85cf
-	data4	0xd06bbbd0, 0xef2ac5ef, 0xaae54faa, 0xfb16edfb
-	data4	0x43c58643, 0x4dd79a4d, 0x33556633, 0x85941185
-	data4	0x45cf8a45, 0xf910e9f9, 0x02060402, 0x7f81fe7f
-	data4	0x50f0a050, 0x3c44783c, 0x9fba259f, 0xa8e34ba8
-	data4	0x51f3a251, 0xa3fe5da3, 0x40c08040, 0x8f8a058f
-	data4	0x92ad3f92, 0x9dbc219d, 0x38487038, 0xf504f1f5
-	data4	0xbcdf63bc, 0xb6c177b6, 0xda75afda, 0x21634221
-	data4	0x10302010, 0xff1ae5ff, 0xf30efdf3, 0xd26dbfd2
-	data4	0xcd4c81cd, 0x0c14180c, 0x13352613, 0xec2fc3ec
-	data4	0x5fe1be5f, 0x97a23597, 0x44cc8844, 0x17392e17
-	data4	0xc45793c4, 0xa7f255a7, 0x7e82fc7e, 0x3d477a3d
-	data4	0x64acc864, 0x5de7ba5d, 0x192b3219, 0x7395e673
-	data4	0x60a0c060, 0x81981981, 0x4fd19e4f, 0xdc7fa3dc
-	data4	0x22664422, 0x2a7e542a, 0x90ab3b90, 0x88830b88
-	data4	0x46ca8c46, 0xee29c7ee, 0xb8d36bb8, 0x143c2814
-	data4	0xde79a7de, 0x5ee2bc5e, 0x0b1d160b, 0xdb76addb
-	data4	0xe03bdbe0, 0x32566432, 0x3a4e743a, 0x0a1e140a
-	data4	0x49db9249, 0x060a0c06, 0x246c4824, 0x5ce4b85c
-	data4	0xc25d9fc2, 0xd36ebdd3, 0xacef43ac, 0x62a6c462
-	data4	0x91a83991, 0x95a43195, 0xe437d3e4, 0x798bf279
-	data4	0xe732d5e7, 0xc8438bc8, 0x37596e37, 0x6db7da6d
-	data4	0x8d8c018d, 0xd564b1d5, 0x4ed29c4e, 0xa9e049a9
-	data4	0x6cb4d86c, 0x56faac56, 0xf407f3f4, 0xea25cfea
-	data4	0x65afca65, 0x7a8ef47a, 0xaee947ae, 0x08181008
-	data4	0xbad56fba, 0x7888f078, 0x256f4a25, 0x2e725c2e
-	data4	0x1c24381c, 0xa6f157a6, 0xb4c773b4, 0xc65197c6
-	data4	0xe823cbe8, 0xdd7ca1dd, 0x749ce874, 0x1f213e1f
-	data4	0x4bdd964b, 0xbddc61bd, 0x8b860d8b, 0x8a850f8a
-	data4	0x7090e070, 0x3e427c3e, 0xb5c471b5, 0x66aacc66
-	data4	0x48d89048, 0x03050603, 0xf601f7f6, 0x0e121c0e
-	data4	0x61a3c261, 0x355f6a35, 0x57f9ae57, 0xb9d069b9
-	data4	0x86911786, 0xc15899c1, 0x1d273a1d, 0x9eb9279e
-	data4	0xe138d9e1, 0xf813ebf8, 0x98b32b98, 0x11332211
-	data4	0x69bbd269, 0xd970a9d9, 0x8e89078e, 0x94a73394
-	data4	0x9bb62d9b, 0x1e223c1e, 0x87921587, 0xe920c9e9
-	data4	0xce4987ce, 0x55ffaa55, 0x28785028, 0xdf7aa5df
-	data4	0x8c8f038c, 0xa1f859a1, 0x89800989, 0x0d171a0d
-	data4	0xbfda65bf, 0xe631d7e6, 0x42c68442, 0x68b8d068
-	data4	0x41c38241, 0x99b02999, 0x2d775a2d, 0x0f111e0f
-	data4	0xb0cb7bb0, 0x54fca854, 0xbbd66dbb, 0x163a2c16
-// Te3:
-	data4	0x6363a5c6, 0x7c7c84f8, 0x777799ee, 0x7b7b8df6
-	data4	0xf2f20dff, 0x6b6bbdd6, 0x6f6fb1de, 0xc5c55491
-	data4	0x30305060, 0x01010302, 0x6767a9ce, 0x2b2b7d56
-	data4	0xfefe19e7, 0xd7d762b5, 0xababe64d, 0x76769aec
-	data4	0xcaca458f, 0x82829d1f, 0xc9c94089, 0x7d7d87fa
-	data4	0xfafa15ef, 0x5959ebb2, 0x4747c98e, 0xf0f00bfb
-	data4	0xadadec41, 0xd4d467b3, 0xa2a2fd5f, 0xafafea45
-	data4	0x9c9cbf23, 0xa4a4f753, 0x727296e4, 0xc0c05b9b
-	data4	0xb7b7c275, 0xfdfd1ce1, 0x9393ae3d, 0x26266a4c
-	data4	0x36365a6c, 0x3f3f417e, 0xf7f702f5, 0xcccc4f83
-	data4	0x34345c68, 0xa5a5f451, 0xe5e534d1, 0xf1f108f9
-	data4	0x717193e2, 0xd8d873ab, 0x31315362, 0x15153f2a
-	data4	0x04040c08, 0xc7c75295, 0x23236546, 0xc3c35e9d
-	data4	0x18182830, 0x9696a137, 0x05050f0a, 0x9a9ab52f
-	data4	0x0707090e, 0x12123624, 0x80809b1b, 0xe2e23ddf
-	data4	0xebeb26cd, 0x2727694e, 0xb2b2cd7f, 0x75759fea
-	data4	0x09091b12, 0x83839e1d, 0x2c2c7458, 0x1a1a2e34
-	data4	0x1b1b2d36, 0x6e6eb2dc, 0x5a5aeeb4, 0xa0a0fb5b
-	data4	0x5252f6a4, 0x3b3b4d76, 0xd6d661b7, 0xb3b3ce7d
-	data4	0x29297b52, 0xe3e33edd, 0x2f2f715e, 0x84849713
-	data4	0x5353f5a6, 0xd1d168b9, 0x00000000, 0xeded2cc1
-	data4	0x20206040, 0xfcfc1fe3, 0xb1b1c879, 0x5b5bedb6
-	data4	0x6a6abed4, 0xcbcb468d, 0xbebed967, 0x39394b72
-	data4	0x4a4ade94, 0x4c4cd498, 0x5858e8b0, 0xcfcf4a85
-	data4	0xd0d06bbb, 0xefef2ac5, 0xaaaae54f, 0xfbfb16ed
-	data4	0x4343c586, 0x4d4dd79a, 0x33335566, 0x85859411
-	data4	0x4545cf8a, 0xf9f910e9, 0x02020604, 0x7f7f81fe
-	data4	0x5050f0a0, 0x3c3c4478, 0x9f9fba25, 0xa8a8e34b
-	data4	0x5151f3a2, 0xa3a3fe5d, 0x4040c080, 0x8f8f8a05
-	data4	0x9292ad3f, 0x9d9dbc21, 0x38384870, 0xf5f504f1
-	data4	0xbcbcdf63, 0xb6b6c177, 0xdada75af, 0x21216342
-	data4	0x10103020, 0xffff1ae5, 0xf3f30efd, 0xd2d26dbf
-	data4	0xcdcd4c81, 0x0c0c1418, 0x13133526, 0xecec2fc3
-	data4	0x5f5fe1be, 0x9