* sysdeps/sparc/Makefile (routines): Reference `dotmul.S', not `mul.S', to avoid a misreference that would override sysdeps/generic/mul.c.

1 files changed, 123 insertions, 0 deletions

diff --git a/sysdeps/sparc/dotmul.S b/sysdeps/sparc/dotmul.S
new file mode 100644
index 0000000000..1c59a27e80
--- /dev/null
+++ b/sysdeps/sparc/dotmul.S
@@ -0,0 +1,123 @@
+/*
+ * Signed multiply, from Appendix E of the Sparc Version 8
+ * Architecture Manual.
+ */
+
+/*
+ * Returns %o0 * %o1 in %o1%o0 (i.e., %o1 holds the upper 32 bits of
+ * the 64-bit product).
+ *
+ * This code optimizes short (less than 13-bit) multiplies.
+ */
+
+#include "DEFS.h"
+FUNC(.mul)
+	mov	%o0, %y		! multiplier -> Y
+	andncc	%o0, 0xfff, %g0	! test bits 12..31
+	be	Lmul_shortway	! if zero, can do it the short way
+	andcc	%g0, %g0, %o4	! zero the partial product and clear N and V
+
+	/*
+	 * Long multiply.  32 steps, followed by a final shift step.
+	 */
+	mulscc	%o4, %o1, %o4	! 1
+	mulscc	%o4, %o1, %o4	! 2
+	mulscc	%o4, %o1, %o4	! 3
+	mulscc	%o4, %o1, %o4	! 4
+	mulscc	%o4, %o1, %o4	! 5
+	mulscc	%o4, %o1, %o4	! 6
+	mulscc	%o4, %o1, %o4	! 7
+	mulscc	%o4, %o1, %o4	! 8
+	mulscc	%o4, %o1, %o4	! 9
+	mulscc	%o4, %o1, %o4	! 10
+	mulscc	%o4, %o1, %o4	! 11
+	mulscc	%o4, %o1, %o4	! 12
+	mulscc	%o4, %o1, %o4	! 13
+	mulscc	%o4, %o1, %o4	! 14
+	mulscc	%o4, %o1, %o4	! 15
+	mulscc	%o4, %o1, %o4	! 16
+	mulscc	%o4, %o1, %o4	! 17
+	mulscc	%o4, %o1, %o4	! 18
+	mulscc	%o4, %o1, %o4	! 19
+	mulscc	%o4, %o1, %o4	! 20
+	mulscc	%o4, %o1, %o4	! 21
+	mulscc	%o4, %o1, %o4	! 22
+	mulscc	%o4, %o1, %o4	! 23
+	mulscc	%o4, %o1, %o4	! 24
+	mulscc	%o4, %o1, %o4	! 25
+	mulscc	%o4, %o1, %o4	! 26
+	mulscc	%o4, %o1, %o4	! 27
+	mulscc	%o4, %o1, %o4	! 28
+	mulscc	%o4, %o1, %o4	! 29
+	mulscc	%o4, %o1, %o4	! 30
+	mulscc	%o4, %o1, %o4	! 31
+	mulscc	%o4, %o1, %o4	! 32
+	mulscc	%o4, %g0, %o4	! final shift
+
+	! If %o0 was negative, the result is
+	!	(%o0 * %o1) + (%o1 << 32))
+	! We fix that here.
+
+#if 0
+	tst	%o0
+	bge	1f
+	rd	%y, %o0
+
+	! %o0 was indeed negative; fix upper 32 bits of result by subtracting 
+	! %o1 (i.e., return %o4 - %o1 in %o1).
+	retl
+	sub	%o4, %o1, %o1
+
+1:
+	retl
+	mov	%o4, %o1
+#else
+	/* Faster code adapted from tege@sics.se's code for umul.S.  */
+	sra	%o0, 31, %o2	! make mask from sign bit
+	and	%o1, %o2, %o2	! %o2 = 0 or %o1, depending on sign of %o0
+	rd	%y, %o0		! get lower half of product
+	retl
+	sub	%o4, %o2, %o1	! subtract compensation 
+				!  and put upper half in place
+#endif
+
+Lmul_shortway:
+	/*
+	 * Short multiply.  12 steps, followed by a final shift step.
+	 * The resulting bits are off by 12 and (32-12) = 20 bit positions,
+	 * but there is no problem with %o0 being negative (unlike above).
+	 */
+	mulscc	%o4, %o1, %o4	! 1
+	mulscc	%o4, %o1, %o4	! 2
+	mulscc	%o4, %o1, %o4	! 3
+	mulscc	%o4, %o1, %o4	! 4
+	mulscc	%o4, %o1, %o4	! 5
+	mulscc	%o4, %o1, %o4	! 6
+	mulscc	%o4, %o1, %o4	! 7
+	mulscc	%o4, %o1, %o4	! 8
+	mulscc	%o4, %o1, %o4	! 9
+	mulscc	%o4, %o1, %o4	! 10
+	mulscc	%o4, %o1, %o4	! 11
+	mulscc	%o4, %o1, %o4	! 12
+	mulscc	%o4, %g0, %o4	! final shift
+
+	/*
+	 *  %o4 has 20 of the bits that should be in the low part of the
+	 * result; %y has the bottom 12 (as %y's top 12).  That is:
+	 *
+	 *	  %o4		    %y
+	 * +----------------+----------------+
+	 * | -12- |   -20-  | -12- |   -20-  |
+	 * +------(---------+------)---------+
+	 *  --hi-- ----low-part----
+	 *
+	 * The upper 12 bits of %o4 should be sign-extended to form the
+	 * high part of the product (i.e., highpart = %o4 >> 20).
+	 */
+
+	rd	%y, %o5
+	sll	%o4, 12, %o0	! shift middle bits left 12
+	srl	%o5, 20, %o5	! shift low bits right 20, zero fill at left
+	or	%o5, %o0, %o0	! construct low part of result
+	retl
+	sra	%o4, 20, %o1	! ... and extract high part of result