- leaq (4*4*4)(%r11), %rax;
- outunpack_blocks(%r11, RA1, RB1, RC1, RD1, RK1, RX0, RY0, RK2);
- outunpack_blocks(%rax, RA2, RB2, RC2, RD2, RK1, RX0, RY0, RK2);
+ outunpack_blocks(RA1, RB1, RC1, RD1, RK1, RX0, RY0, RK2);
+ outunpack_blocks(RA2, RB2, RC2, RD2, RK1, RX0, RY0, RK2);
+
+ ret;
+
+.align 8
+.global twofish_ecb_enc_8way
+.type twofish_ecb_enc_8way,@function;
+
+twofish_ecb_enc_8way:
+ /* input:
+ * %rdi: ctx, CTX
+ * %rsi: dst
+ * %rdx: src
+ */
+
+ movq %rsi, %r11;
+
+ load_8way(%rdx, RA1, RB1, RC1, RD1, RA2, RB2, RC2, RD2);
+
+ call __twofish_enc_blk8;
+
+ store_8way(%r11, RC1, RD1, RA1, RB1, RC2, RD2, RA2, RB2);
+
+ ret;
+
+.align 8
+.global twofish_ecb_dec_8way
+.type twofish_ecb_dec_8way,@function;
+
+twofish_ecb_dec_8way:
+ /* input:
+ * %rdi: ctx, CTX
+ * %rsi: dst
+ * %rdx: src
+ */
+
+ movq %rsi, %r11;
+
+ load_8way(%rdx, RC1, RD1, RA1, RB1, RC2, RD2, RA2, RB2);
+
+ call __twofish_dec_blk8;
+
+ store_8way(%r11, RA1, RB1, RC1, RD1, RA2, RB2, RC2, RD2);
+
+ ret;
+
+.align 8
+.global twofish_cbc_dec_8way
+.type twofish_cbc_dec_8way,@function;
+
+twofish_cbc_dec_8way:
+ /* input:
+ * %rdi: ctx, CTX
+ * %rsi: dst
+ * %rdx: src
+ */
+
+ pushq %r12;
+
+ movq %rsi, %r11;
+ movq %rdx, %r12;
+
+ load_8way(%rdx, RC1, RD1, RA1, RB1, RC2, RD2, RA2, RB2);
+
+ call __twofish_dec_blk8;
+
+ store_cbc_8way(%r12, %r11, RA1, RB1, RC1, RD1, RA2, RB2, RC2, RD2);
+
+ popq %r12;
+
+ ret;
+
+.align 8
+.global twofish_ctr_8way
+.type twofish_ctr_8way,@function;
+
+twofish_ctr_8way:
+ /* input:
+ * %rdi: ctx, CTX
+ * %rsi: dst
+ * %rdx: src
+ * %rcx: iv (little endian, 128bit)
+ */
+
+ pushq %r12;
+
+ movq %rsi, %r11;
+ movq %rdx, %r12;
+
+ load_ctr_8way(%rcx, .Lbswap128_mask, RA1, RB1, RC1, RD1, RA2, RB2, RC2,
+ RD2, RX0, RX1, RY0);
+
+ call __twofish_enc_blk8;
+
+ store_ctr_8way(%r12, %r11, RC1, RD1, RA1, RB1, RC2, RD2, RA2, RB2);
+
+ popq %r12;