Import of old SSLeay release: SSLeay 0.9.0b

4 files changed, 2124 insertions, 0 deletions

diff --git a/crypto/cast/asm/c-win32.asm b/crypto/cast/asm/c-win32.asm
new file mode 100644
index 0000000000..a1d8a2671a
--- /dev/null
+++ b/crypto/cast/asm/c-win32.asm
@@ -0,0 +1,940 @@
+	; Don't even think of reading this code
+	; It was automatically generated by cast-586.pl
+	; Which is a perl program used to generate the x86 assember for
+	; any of elf, a.out, BSDI,Win32, or Solaris
+	; eric <eay@cryptsoft.com>
+	; 
+	TITLE	cast-586.asm
+        .486
+.model FLAT
+_TEXT	SEGMENT
+PUBLIC	_CAST_encrypt
+EXTERN	_CAST_S_table0:DWORD
+EXTERN	_CAST_S_table1:DWORD
+EXTERN	_CAST_S_table2:DWORD
+EXTERN	_CAST_S_table3:DWORD
+
+_CAST_encrypt PROC NEAR
+	; 
+	push	ebp
+	push	ebx
+	mov	ebx,		DWORD PTR 12[esp]
+	mov	ebp,		DWORD PTR 16[esp]
+	push	esi
+	push	edi
+	; Load the 2 words
+	mov	edi,		DWORD PTR [ebx]
+	mov	esi,		DWORD PTR 4[ebx]
+	xor	eax,		eax
+	; round 0
+	mov	edx,		DWORD PTR [ebp]
+	mov	ecx,		DWORD PTR 4[ebp]
+	add	edx,		esi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,		ebx
+	xor	edi,		ecx
+	; round 1
+	mov	edx,		DWORD PTR 8[ebp]
+	mov	ecx,		DWORD PTR 12[ebp]
+	xor	edx,		edi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,		ebx
+	xor	esi,		ecx
+	; round 2
+	mov	edx,		DWORD PTR 16[ebp]
+	mov	ecx,		DWORD PTR 20[ebp]
+	sub	edx,		esi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,		ebx
+	xor	edi,		ecx
+	; round 3
+	mov	edx,		DWORD PTR 24[ebp]
+	mov	ecx,		DWORD PTR 28[ebp]
+	add	edx,		edi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,		ebx
+	xor	esi,		ecx
+	; round 4
+	mov	edx,		DWORD PTR 32[ebp]
+	mov	ecx,		DWORD PTR 36[ebp]
+	xor	edx,		esi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,		ebx
+	xor	edi,		ecx
+	; round 5
+	mov	edx,		DWORD PTR 40[ebp]
+	mov	ecx,		DWORD PTR 44[ebp]
+	sub	edx,		edi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,		ebx
+	xor	esi,		ecx
+	; round 6
+	mov	edx,		DWORD PTR 48[ebp]
+	mov	ecx,		DWORD PTR 52[ebp]
+	add	edx,		esi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,		ebx
+	xor	edi,		ecx
+	; round 7
+	mov	edx,		DWORD PTR 56[ebp]
+	mov	ecx,		DWORD PTR 60[ebp]
+	xor	edx,		edi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,		ebx
+	xor	esi,		ecx
+	; round 8
+	mov	edx,		DWORD PTR 64[ebp]
+	mov	ecx,		DWORD PTR 68[ebp]
+	sub	edx,		esi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,		ebx
+	xor	edi,		ecx
+	; round 9
+	mov	edx,		DWORD PTR 72[ebp]
+	mov	ecx,		DWORD PTR 76[ebp]
+	add	edx,		edi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,		ebx
+	xor	esi,		ecx
+	; round 10
+	mov	edx,		DWORD PTR 80[ebp]
+	mov	ecx,		DWORD PTR 84[ebp]
+	xor	edx,		esi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,		ebx
+	xor	edi,		ecx
+	; round 11
+	mov	edx,		DWORD PTR 88[ebp]
+	mov	ecx,		DWORD PTR 92[ebp]
+	sub	edx,		edi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,		ebx
+	xor	esi,		ecx
+	; round 12
+	mov	edx,		DWORD PTR 96[ebp]
+	mov	ecx,		DWORD PTR 100[ebp]
+	add	edx,		esi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,		ebx
+	xor	edi,		ecx
+	; round 13
+	mov	edx,		DWORD PTR 104[ebp]
+	mov	ecx,		DWORD PTR 108[ebp]
+	xor	edx,		edi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,		ebx
+	xor	esi,		ecx
+	; round 14
+	mov	edx,		DWORD PTR 112[ebp]
+	mov	ecx,		DWORD PTR 116[ebp]
+	sub	edx,		esi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,		ebx
+	xor	edi,		ecx
+	; round 15
+	mov	edx,		DWORD PTR 120[ebp]
+	mov	ecx,		DWORD PTR 124[ebp]
+	add	edx,		edi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,		ebx
+	mov	eax,		DWORD PTR 20[esp]
+	xor	esi,		ecx
+	nop
+	mov	DWORD PTR 4[eax],edi
+	mov	DWORD PTR [eax],esi
+	pop	edi
+	pop	esi
+	pop	ebx
+	pop	ebp
+	ret
+_CAST_encrypt ENDP
+_TEXT	ENDS
+_TEXT	SEGMENT
+PUBLIC	_CAST_decrypt
+EXTERN	_CAST_S_table0:DWORD
+EXTERN	_CAST_S_table1:DWORD
+EXTERN	_CAST_S_table2:DWORD
+EXTERN	_CAST_S_table3:DWORD
+
+_CAST_decrypt PROC NEAR
+	; 
+	push	ebp
+	push	ebx
+	mov	ebx,		DWORD PTR 12[esp]
+	mov	ebp,		DWORD PTR 16[esp]
+	push	esi
+	push	edi
+	; Load the 2 words
+	mov	edi,		DWORD PTR [ebx]
+	mov	esi,		DWORD PTR 4[ebx]
+	xor	eax,		eax
+	; round 15
+	mov	edx,		DWORD PTR 120[ebp]
+	mov	ecx,		DWORD PTR 124[ebp]
+	add	edx,		esi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,		ebx
+	xor	edi,		ecx
+	; round 14
+	mov	edx,		DWORD PTR 112[ebp]
+	mov	ecx,		DWORD PTR 116[ebp]
+	sub	edx,		edi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,		ebx
+	xor	esi,		ecx
+	; round 13
+	mov	edx,		DWORD PTR 104[ebp]
+	mov	ecx,		DWORD PTR 108[ebp]
+	xor	edx,		esi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,		ebx
+	xor	edi,		ecx
+	; round 12
+	mov	edx,		DWORD PTR 96[ebp]
+	mov	ecx,		DWORD PTR 100[ebp]
+	add	edx,		edi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,		ebx
+	xor	esi,		ecx
+	; round 11
+	mov	edx,		DWORD PTR 88[ebp]
+	mov	ecx,		DWORD PTR 92[ebp]
+	sub	edx,		esi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,		ebx
+	xor	edi,		ecx
+	; round 10
+	mov	edx,		DWORD PTR 80[ebp]
+	mov	ecx,		DWORD PTR 84[ebp]
+	xor	edx,		edi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,		ebx
+	xor	esi,		ecx
+	; round 9
+	mov	edx,		DWORD PTR 72[ebp]
+	mov	ecx,		DWORD PTR 76[ebp]
+	add	edx,		esi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,		ebx
+	xor	edi,		ecx
+	; round 8
+	mov	edx,		DWORD PTR 64[ebp]
+	mov	ecx,		DWORD PTR 68[ebp]
+	sub	edx,		edi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,		ebx
+	xor	esi,		ecx
+	; round 7
+	mov	edx,		DWORD PTR 56[ebp]
+	mov	ecx,		DWORD PTR 60[ebp]
+	xor	edx,		esi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,		ebx
+	xor	edi,		ecx
+	; round 6
+	mov	edx,		DWORD PTR 48[ebp]
+	mov	ecx,		DWORD PTR 52[ebp]
+	add	edx,		edi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,		ebx
+	xor	esi,		ecx
+	; round 5
+	mov	edx,		DWORD PTR 40[ebp]
+	mov	ecx,		DWORD PTR 44[ebp]
+	sub	edx,		esi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,		ebx
+	xor	edi,		ecx
+	; round 4
+	mov	edx,		DWORD PTR 32[ebp]
+	mov	ecx,		DWORD PTR 36[ebp]
+	xor	edx,		edi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,		ebx
+	xor	esi,		ecx
+	; round 3
+	mov	edx,		DWORD PTR 24[ebp]
+	mov	ecx,		DWORD PTR 28[ebp]
+	add	edx,		esi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,		ebx
+	xor	edi,		ecx
+	; round 2
+	mov	edx,		DWORD PTR 16[ebp]
+	mov	ecx,		DWORD PTR 20[ebp]
+	sub	edx,		edi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,		ebx
+	xor	esi,		ecx
+	; round 1
+	mov	edx,		DWORD PTR 8[ebp]
+	mov	ecx,		DWORD PTR 12[ebp]
+	xor	edx,		esi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,		ebx
+	xor	edi,		ecx
+	; round 0
+	mov	edx,		DWORD PTR [ebp]
+	mov	ecx,		DWORD PTR 4[ebp]
+	add	edx,		edi
+	rol	edx,		cl
+	mov	ebx,		edx
+	xor	ecx,		ecx
+	mov	cl,		dh
+	and	ebx,		255
+	shr	edx,		16
+	xor	eax,		eax
+	mov	al,		dh
+	and	edx,		255
+	mov	ecx,		DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,		DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,		ebx
+	mov	ebx,		DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,		ebx
+	mov	eax,		DWORD PTR 20[esp]
+	xor	esi,		ecx
+	nop
+	mov	DWORD PTR 4[eax],edi
+	mov	DWORD PTR [eax],esi
+	pop	edi
+	pop	esi
+	pop	ebx
+	pop	ebp
+	ret
+_CAST_decrypt ENDP
+_TEXT	ENDS
+_TEXT	SEGMENT
+PUBLIC	_CAST_cbc_encrypt
+
+_CAST_cbc_encrypt PROC NEAR
+	; 
+	push	ebp
+	push	ebx
+	push	esi
+	push	edi
+	mov	ebp,		DWORD PTR 28[esp]
+	; getting iv ptr from parameter 4
+	mov	ebx,		DWORD PTR 36[esp]
+	mov	esi,		DWORD PTR [ebx]
+	mov	edi,		DWORD PTR 4[ebx]
+	push	edi
+	push	esi
+	push	edi
+	push	esi
+	mov	ebx,		esp
+	mov	esi,		DWORD PTR 36[esp]
+	mov	edi,		DWORD PTR 40[esp]
+	; getting encrypt flag from parameter 5
+	mov	ecx,		DWORD PTR 56[esp]
+	; get and push parameter 3
+	mov	eax,		DWORD PTR 48[esp]
+	push	eax
+	push	ebx
+	cmp	ecx,		0
+	jz	$L000decrypt
+	and	ebp,		4294967288
+	mov	eax,		DWORD PTR 8[esp]
+	mov	ebx,		DWORD PTR 12[esp]
+	jz	$L001encrypt_finish
+L002encrypt_loop:
+	mov	ecx,		DWORD PTR [esi]
+	mov	edx,		DWORD PTR 4[esi]
+	xor	eax,		ecx
+	xor	ebx,		edx
+	bswap	eax
+	bswap	ebx
+	mov	DWORD PTR 8[esp],eax
+	mov	DWORD PTR 12[esp],ebx
+	call	_CAST_encrypt
+	mov	eax,		DWORD PTR 8[esp]
+	mov	ebx,		DWORD PTR 12[esp]
+	bswap	eax
+	bswap	ebx
+	mov	DWORD PTR [edi],eax
+	mov	DWORD PTR 4[edi],ebx
+	add	esi,		8
+	add	edi,		8
+	sub	ebp,		8
+	jnz	L002encrypt_loop
+$L001encrypt_finish:
+	mov	ebp,		DWORD PTR 52[esp]
+	and	ebp,		7
+	jz	$L003finish
+	xor	ecx,		ecx
+	xor	edx,		edx
+	mov	ebp,		DWORD PTR $L004cbc_enc_jmp_table[ebp*4]
+	jmp	 ebp
+L005ej7:
+	xor	edx,		edx
+	mov	dh,		BYTE PTR 6[esi]
+	shl	edx,		8
+L006ej6:
+	mov	dh,		BYTE PTR 5[esi]
+L007ej5:
+	mov	dl,		BYTE PTR 4[esi]
+L008ej4:
+	mov	ecx,		DWORD PTR [esi]
+	jmp	$L009ejend
+L010ej3:
+	mov	ch,		BYTE PTR 2[esi]
+	xor	ecx,		ecx
+	shl	ecx,		8
+L011ej2:
+	mov	ch,		BYTE PTR 1[esi]
+L012ej1:
+	mov	cl,		BYTE PTR [esi]
+$L009ejend:
+	xor	eax,		ecx
+	xor	ebx,		edx
+	bswap	eax
+	bswap	ebx
+	mov	DWORD PTR 8[esp],eax
+	mov	DWORD PTR 12[esp],ebx
+	call	_CAST_encrypt
+	mov	eax,		DWORD PTR 8[esp]
+	mov	ebx,		DWORD PTR 12[esp]
+	bswap	eax
+	bswap	ebx
+	mov	DWORD PTR [edi],eax
+	mov	DWORD PTR 4[edi],ebx
+	jmp	$L003finish
+$L000decrypt:
+	and	ebp,		4294967288
+	mov	eax,		DWORD PTR 16[esp]
+	mov	ebx,		DWORD PTR 20[esp]
+	jz	$L013decrypt_finish
+L014decrypt_loop:
+	mov	eax,		DWORD PTR [esi]
+	mov	ebx,		DWORD PTR 4[esi]
+	bswap	eax
+	bswap	ebx
+	mov	DWORD PTR 8[esp],eax
+	mov	DWORD PTR 12[esp],ebx
+	call	_CAST_decrypt
+	mov	eax,		DWORD PTR 8[esp]
+	mov	ebx,		DWORD PTR 12[esp]
+	bswap	eax
+	bswap	ebx
+	mov	ecx,		DWORD PTR 16[esp]
+	mov	edx,		DWORD PTR 20[esp]
+	xor	ecx,		eax
+	xor	edx,		ebx
+	mov	eax,		DWORD PTR [esi]
+	mov	ebx,		DWORD PTR 4[esi]
+	mov	DWORD PTR [edi],ecx
+	mov	DWORD PTR 4[edi],edx
+	mov	DWORD PTR 16[esp],eax
+	mov	DWORD PTR 20[esp],ebx
+	add	esi,		8
+	add	edi,		8
+	sub	ebp,		8
+	jnz	L014decrypt_loop
+$L013decrypt_finish:
+	mov	ebp,		DWORD PTR 52[esp]
+	and	ebp,		7
+	jz	$L003finish
+	mov	eax,		DWORD PTR [esi]
+	mov	ebx,		DWORD PTR 4[esi]
+	bswap	eax
+	bswap	ebx
+	mov	DWORD PTR 8[esp],eax
+	mov	DWORD PTR 12[esp],ebx
+	call	_CAST_decrypt
+	mov	eax,		DWORD PTR 8[esp]
+	mov	ebx,		DWORD PTR 12[esp]
+	bswap	eax
+	bswap	ebx
+	mov	ecx,		DWORD PTR 16[esp]
+	mov	edx,		DWORD PTR 20[esp]
+	xor	ecx,		eax
+	xor	edx,		ebx
+	mov	eax,		DWORD PTR [esi]
+	mov	ebx,		DWORD PTR 4[esi]
+L015dj7:
+	ror	edx,		16
+	mov	BYTE PTR 6[edi],dl
+	shr	edx,		16
+L016dj6:
+	mov	BYTE PTR 5[edi],dh
+L017dj5:
+	mov	BYTE PTR 4[edi],dl
+L018dj4:
+	mov	DWORD PTR [edi],ecx
+	jmp	$L019djend
+L020dj3:
+	ror	ecx,		16
+	mov	BYTE PTR 2[edi],cl
+	shl	ecx,		16
+L021dj2:
+	mov	BYTE PTR 1[esi],ch
+L022dj1:
+	mov	BYTE PTR [esi],	cl
+$L019djend:
+	jmp	$L003finish
+$L003finish:
+	mov	ecx,		DWORD PTR 60[esp]
+	add	esp,		24
+	mov	DWORD PTR [ecx],eax
+	mov	DWORD PTR 4[ecx],ebx
+	pop	edi
+	pop	esi
+	pop	ebx
+	pop	ebp
+	ret
+$L004cbc_enc_jmp_table:
+	DD	0
+	DD	L012ej1
+	DD	L011ej2
+	DD	L010ej3
+	DD	L008ej4
+	DD	L007ej5
+	DD	L006ej6
+	DD	L005ej7
+L023cbc_dec_jmp_table:
+	DD	0
+	DD	L022dj1
+	DD	L021dj2
+	DD	L020dj3
+	DD	L018dj4
+	DD	L017dj5
+	DD	L016dj6
+	DD	L015dj7
+_CAST_cbc_encrypt ENDP
+_TEXT	ENDS
+END
diff --git a/crypto/cast/asm/cast-586.pl b/crypto/cast/asm/cast-586.pl
new file mode 100644
index 0000000000..d6b6f19bea
--- /dev/null
+++ b/crypto/cast/asm/cast-586.pl
@@ -0,0 +1,167 @@
+#!/usr/local/bin/perl
+
+# define for pentium pro friendly version
+$ppro=1;
+
+push(@INC,"perlasm","../../perlasm");
+require "x86asm.pl";
+require "cbc.pl";
+
+&asm_init($ARGV[0],"cast-586.pl");
+
+$CAST_ROUNDS=16;
+$L="edi";
+$R="esi";
+$K="ebp";
+$tmp1="ecx";
+$tmp2="ebx";
+$tmp3="eax";
+$tmp4="edx";
+$S1="CAST_S_table0";
+$S2="CAST_S_table1";
+$S3="CAST_S_table2";
+$S4="CAST_S_table3";
+
+@F1=("add","xor","sub");
+@F2=("xor","sub","add");
+@F3=("sub","add","xor");
+
+&CAST_encrypt("CAST_encrypt",1);
+&CAST_encrypt("CAST_decrypt",0);
+&cbc("CAST_cbc_encrypt","CAST_encrypt","CAST_decrypt",1,4,5,3,-1,-1);
+
+&asm_finish();
+
+sub CAST_encrypt
+	{
+	local($name,$enc)=@_;
+
+	local($win_ex)=<<"EOF";
+EXTERN	_CAST_S_table0:DWORD
+EXTERN	_CAST_S_table1:DWORD
+EXTERN	_CAST_S_table2:DWORD
+EXTERN	_CAST_S_table3:DWORD
+EOF
+	&main'external_label(
+		"CAST_S_table0",
+		"CAST_S_table1",
+		"CAST_S_table2",
+		"CAST_S_table3",
+		);
+
+	&function_begin_B($name,$win_ex);
+
+	&comment("");
+
+	&push("ebp");
+	&push("ebx");
+	&mov($tmp2,&wparam(0));
+	&mov($K,&wparam(1));
+	&push("esi");
+	&push("edi");
+
+	&comment("Load the 2 words");
+	&mov($L,&DWP(0,$tmp2,"",0));
+	&mov($R,&DWP(4,$tmp2,"",0));
+
+	&xor(	$tmp3,	$tmp3);
+
+	# encrypting part
+
+	if ($enc)
+		{
+		&E_CAST( 0,$S,$L,$R,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 1,$S,$R,$L,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 2,$S,$L,$R,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 3,$S,$R,$L,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 4,$S,$L,$R,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 5,$S,$R,$L,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 6,$S,$L,$R,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 7,$S,$R,$L,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 8,$S,$L,$R,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 9,$S,$R,$L,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST(10,$S,$L,$R,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST(11,$S,$R,$L,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST(12,$S,$L,$R,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST(13,$S,$R,$L,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST(14,$S,$L,$R,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST(15,$S,$R,$L,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4,1);
+		}
+	else
+		{
+		&E_CAST(15,$S,$L,$R,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST(14,$S,$R,$L,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST(13,$S,$L,$R,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST(12,$S,$R,$L,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST(11,$S,$L,$R,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST(10,$S,$R,$L,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 9,$S,$L,$R,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 8,$S,$R,$L,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 7,$S,$L,$R,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 6,$S,$R,$L,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 5,$S,$L,$R,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 4,$S,$R,$L,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 3,$S,$L,$R,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 2,$S,$R,$L,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 1,$S,$L,$R,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
+		&E_CAST( 0,$S,$R,$L,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4,1);
+		}
+
+	&nop();
+	&mov(&DWP(4,$tmp3,"",0),$L);
+	&mov(&DWP(0,$tmp3,"",0),$R);
+	&function_end($name);
+	}
+
+sub E_CAST
+	{
+	local($i,$S,$L,$R,$K,$OP1,$OP2,$OP3,$tmp1,$tmp2,$tmp3,$tmp4,$lst)=@_;
+	# Ri needs to have 16 pre added.
+
+	&comment("round $i");
+	&mov(	$tmp4,		&DWP($i*8,$K,"",1));
+
+	&mov(	$tmp1,		&DWP($i*8+4,$K,"",1));# must be word
+	&$OP1(	$tmp4,		$R);
+
+	&rotl(	$tmp4,		&LB($tmp1));
+
+	if ($ppro)
+		{
+		&mov(	$tmp2,		$tmp4);		# B
+		&xor(	$tmp1,		$tmp1);
+
+		&movb(	&LB($tmp1),	&HB($tmp4));	# A
+		&and(	$tmp2,		0xff);
+
+		&shr(	$tmp4,		16); 		#
+		&xor(	$tmp3,		$tmp3);
+		}
+	else
+		{
+		&mov(	$tmp2,		$tmp4);		# B
+		&movb(	&LB($tmp1),	&HB($tmp4));	# A	# BAD BAD BAD
+
+		&shr(	$tmp4,		16); 		#
+		&and(	$tmp2,		0xff);
+		}
+
+	&movb(	&LB($tmp3),	&HB($tmp4));	# C	# BAD BAD BAD
+	&and(	$tmp4,		0xff);		# D
+
+	&mov(	$tmp1,		&DWP($S1,"",$tmp1,4));
+	&mov(	$tmp2,		&DWP($S2,"",$tmp2,4));
+
+	&$OP2(	$tmp1,		$tmp2);
+	&mov(	$tmp2,		&DWP($S3,"",$tmp3,4));
+
+	&$OP3(	$tmp1,		$tmp2);
+	&mov(	$tmp2,		&DWP($S4,"",$tmp4,4));
+
+	&$OP1(	$tmp1,		$tmp2);
+	 &mov($tmp3,&wparam(0)) if $lst;
+	 # XXX
+
+	&xor(	$L,		$tmp1);
+	 # XXX
+	}
diff --git a/crypto/cast/asm/cx86unix.cpp b/crypto/cast/asm/cx86unix.cpp
new file mode 100644
index 0000000000..035692a5af
--- /dev/null
+++ b/crypto/cast/asm/cx86unix.cpp
@@ -0,0 +1,1010 @@
+/* Run the C pre-processor over this file with one of the following defined
+ * ELF - elf object files,
+ * OUT - a.out object files,
+ * BSDI - BSDI style a.out object files
+ * SOL - Solaris style elf
+ */
+
+#define TYPE(a,b)       .type   a,b
+#define SIZE(a,b)       .size   a,b
+
+#if defined(OUT) || defined(BSDI)
+#define CAST_S_table0 _CAST_S_table0
+#define CAST_S_table1 _CAST_S_table1
+#define CAST_S_table2 _CAST_S_table2
+#define CAST_S_table3 _CAST_S_table3
+#define CAST_encrypt _CAST_encrypt
+#define CAST_S_table0 _CAST_S_table0
+#define CAST_S_table1 _CAST_S_table1
+#define CAST_S_table2 _CAST_S_table2
+#define CAST_S_table3 _CAST_S_table3
+#define CAST_decrypt _CAST_decrypt
+#define CAST_cbc_encrypt _CAST_cbc_encrypt
+
+#endif
+
+#ifdef OUT
+#define OK	1
+#define ALIGN	4
+#endif
+
+#ifdef BSDI
+#define OK              1
+#define ALIGN           4
+#undef SIZE
+#undef TYPE
+#define SIZE(a,b)
+#define TYPE(a,b)
+#endif
+
+#if defined(ELF) || defined(SOL)
+#define OK              1
+#define ALIGN           16
+#endif
+
+#ifndef OK
+You need to define one of
+ELF - elf systems - linux-elf, NetBSD and DG-UX
+OUT - a.out systems - linux-a.out and FreeBSD
+SOL - solaris systems, which are elf with strange comment lines
+BSDI - a.out with a very primative version of as.
+#endif
+
+/* Let the Assembler begin :-) */
+	/* Don't even think of reading this code */
+	/* It was automatically generated by cast-586.pl */
+	/* Which is a perl program used to generate the x86 assember for */
+	/* any of elf, a.out, BSDI,Win32, or Solaris */
+	/* eric <eay@cryptsoft.com> */
+
+	.file	"cast-586.s"
+	.version	"01.01"
+gcc2_compiled.:
+.text
+	.align ALIGN
+.globl CAST_encrypt
+	TYPE(CAST_encrypt,@function)
+CAST_encrypt:
+
+	pushl	%ebp
+	pushl	%ebx
+	movl	12(%esp),	%ebx
+	movl	16(%esp),	%ebp
+	pushl	%esi
+	pushl	%edi
+	/* Load the 2 words */
+	movl	(%ebx),		%edi
+	movl	4(%ebx),	%esi
+	xorl	%eax,		%eax
+	/* round 0 */
+	movl	(%ebp),		%edx
+	movl	4(%ebp),	%ecx
+	addl	%esi,		%edx
+	roll	%cl,		%edx
+	movl	%edx,		%ebx
+	xorl	%ecx,		%ecx
+	movb	%dh,		%cl
+	andl	$255,		%ebx
+	shrl	$16,		%edx
+	xorl	%eax,		%eax
+	movb	%dh,		%al
+	andl	$255,		%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,		%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,		%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,		%ecx
+	xorl	%ecx,		%edi
+	/* round 1 */
+	movl	8(%ebp),	%edx
+	movl	12(%ebp),	%ecx
+	xorl	%edi,		%edx
+	roll	%cl,		%edx
+	movl	%edx,		%ebx
+	xorl	%ecx,		%ecx
+	movb	%dh,		%cl
+	andl	$255,		%ebx
+	shrl	$16,		%edx
+	xorl	%eax,		%eax
+	movb	%dh,		%al
+	andl	$255,		%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,		%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx