Blackfin arch: Replace C version of 64 bit multiply with hand optimiz…

…ed assembly

Signed-off-by: Bernd Schmidt <bernds_cb1@t-online.de>
Signed-off-by: Bryan Wu <cooloney@kernel.org>

arch/blackfin/lib/muldi3.S

@@ -0,0 +1,68 @@
+.align 2
+.global ___muldi3;
+.type ___muldi3, STT_FUNC;
+
+#ifdef CONFIG_ARITHMETIC_OPS_L1
+.section .l1.text
+#else
+.text
+#endif
+
+/*
+	   R1:R0 * R3:R2
+	 = R1.h:R1.l:R0.h:R0.l * R3.h:R3.l:R2.h:R2.l
+[X]	 = (R1.h * R3.h) * 2^96
+[X]	   + (R1.h * R3.l + R1.l * R3.h) * 2^80
+[X]	   + (R1.h * R2.h + R1.l * R3.l + R3.h * R0.h) * 2^64
+[T1]	   + (R1.h * R2.l + R3.h * R0.l + R1.l * R2.h + R3.l * R0.h) * 2^48
+[T2]	   + (R1.l * R2.l + R3.l * R0.l + R0.h * R2.h) * 2^32
+[T3]	   + (R0.l * R2.h + R2.l * R0.h) * 2^16
+[T4]	   + (R0.l * R2.l)
+
+	We can discard the first three lines marked "X" since we produce
+	only a 64 bit result.  So, we need ten 16-bit multiplies.
+
+	Individual mul-acc results:
+[E1]	 =  R1.h * R2.l + R3.h * R0.l + R1.l * R2.h + R3.l * R0.h
+[E2]	 =  R1.l * R2.l + R3.l * R0.l + R0.h * R2.h
+[E3]	 =  R0.l * R2.h + R2.l * R0.h
+[E4]	 =  R0.l * R2.l
+
+	We also need to add high parts from lower-level results to higher ones:
+	E[n]c = E[n] + (E[n+1]c >> 16), where E4c := E4
+
+	One interesting property is that all parts of the result that depend
+	on the sign of the multiplication are discarded.  Those would be the
+	multiplications involving R1.h and R3.h, but only the top 16 bit of
+	the 32 bit result depend on the sign, and since R1.h and R3.h only
+	occur in E1, the top half of these results is cut off.
+	So, we can just use FU mode for all of the 16-bit multiplies, and
+	ignore questions of when to use mixed mode.  */
+
+___muldi3:
+	/* [SP] technically is part of the caller's frame, but we can
+	   use it as scratch space.  */
+	A0 = R2.H * R1.L, A1 = R2.L * R1.H (FU) || R3 = [SP + 12];	/* E1 */
+	A0 += R3.H * R0.L, A1 += R3.L * R0.H (FU) || [SP] = R4;		/* E1 */
+	A0 += A1;							/* E1 */
+	R4 = A0.w;
+	A0 = R0.l * R3.l (FU);						/* E2 */
+	A0 += R2.l * R1.l (FU);						/* E2 */
+
+	A1 = R2.L * R0.L (FU);						/* E4 */
+	R3 = A1.w;
+	A1 = A1 >> 16;							/* E3c */
+	A0 += R2.H * R0.H, A1 += R2.L * R0.H (FU);			/* E2, E3c */
+	A1 += R0.L * R2.H (FU);						/* E3c */
+	R0 = A1.w;
+	A1 = A1 >> 16;							/* E2c */
+	A0 += A1;							/* E2c */
+	R1 = A0.w;
+
+	/* low(result) = low(E3c):low(E4) */
+	R0 = PACK (R0.l, R3.l);
+	/* high(result) = E2c + (E1 << 16) */
+	R1.h = R1.h + R4.l (NS) || R4 = [SP];
+	RTS;
+
+.size ___muldi3, .-___muldi3