Basic constant folding implementation

3 files changed, 291 insertions, 0 deletions

diff --git a/vendor/gmp-6.3.0/mpn/alpha/ev67/gcd_11.asm b/vendor/gmp-6.3.0/mpn/alpha/ev67/gcd_11.asm
new file mode 100644
index 0000000..03c234b
--- /dev/null
+++ b/vendor/gmp-6.3.0/mpn/alpha/ev67/gcd_11.asm
@@ -0,0 +1,79 @@
+dnl  Alpha ev67 mpn_gcd_11 -- Nx1 greatest common divisor.
+
+dnl  Copyright 2003, 2004 Free Software Foundation, Inc.
+
+dnl  This file is part of the GNU MP Library.
+dnl
+dnl  The GNU MP Library is free software; you can redistribute it and/or modify
+dnl  it under the terms of either:
+dnl
+dnl    * the GNU Lesser General Public License as published by the Free
+dnl      Software Foundation; either version 3 of the License, or (at your
+dnl      option) any later version.
+dnl
+dnl  or
+dnl
+dnl    * the GNU General Public License as published by the Free Software
+dnl      Foundation; either version 2 of the License, or (at your option) any
+dnl      later version.
+dnl
+dnl  or both in parallel, as here.
+dnl
+dnl  The GNU MP Library is distributed in the hope that it will be useful, but
+dnl  WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
+dnl  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
+dnl  for more details.
+dnl
+dnl  You should have received copies of the GNU General Public License and the
+dnl  GNU Lesser General Public License along with the GNU MP Library.  If not,
+dnl  see https://www.gnu.org/licenses/.
+
+include(`../config.m4')
+
+
+C ev67: 3.4 cycles/bitpair for 1x1 part
+
+
+C mp_limb_t mpn_gcd_1 (mp_srcptr xp, mp_size_t xsize, mp_limb_t y);
+C
+C In the 1x1 part, the algorithm is to change x,y to abs(x-y),min(x,y) and
+C strip trailing zeros from abs(x-y) to maintain x and y both odd.
+C
+C The trailing zeros are calculated from just x-y, since in twos-complement
+C there's the same number of trailing zeros on d or -d.  This means the cttz
+C runs in parallel with abs(x-y).
+C
+C The loop takes 5 cycles, and at 0.68 iterations per bit for two N-bit
+C operands with this algorithm gives the measured 3.4 c/l.
+C
+C The slottings shown are for SVR4 style systems, Unicos differs in the
+C initial gp setup and the LEA.
+
+
+ASM_START()
+PROLOGUE(mpn_gcd_11)
+	mov	r16, r0
+	mov	r17, r1
+
+	ALIGN(16)
+L(top):	subq	r0, r1, r7		C l0  d = x - y
+	cmpult	r0, r1, r16		C u0  test x >= y
+
+	subq	r1, r0, r4		C l0  new_x = y - x
+	cttz	r7, r8			C U0  d twos
+
+	cmoveq	r16, r7, r4		C l0  new_x = d if x>=y
+	cmovne	r16, r0, r1		C u0  y = x if x<y
+	unop				C l   \ force cmoveq into l0
+	unop				C u   /
+
+	C				C cmoveq2 L0, cmovne2 U0
+
+	srl	r4, r8, r0		C U0  x = new_x >> twos
+	bne	r7, L(top)		C U1  stop when d==0
+
+
+L(end):	mov	r1, r0			C U0  return y << common_twos
+	ret	r31, (r26), 1		C L0
+EPILOGUE()
+ASM_END()
diff --git a/vendor/gmp-6.3.0/mpn/alpha/ev67/hamdist.asm b/vendor/gmp-6.3.0/mpn/alpha/ev67/hamdist.asm
new file mode 100644
index 0000000..4b13e9f
--- /dev/null
+++ b/vendor/gmp-6.3.0/mpn/alpha/ev67/hamdist.asm
@@ -0,0 +1,111 @@
+dnl  Alpha ev67 mpn_hamdist -- mpn hamming distance.
+
+dnl  Copyright 2003, 2005 Free Software Foundation, Inc.
+
+dnl  This file is part of the GNU MP Library.
+dnl
+dnl  The GNU MP Library is free software; you can redistribute it and/or modify
+dnl  it under the terms of either:
+dnl
+dnl    * the GNU Lesser General Public License as published by the Free
+dnl      Software Foundation; either version 3 of the License, or (at your
+dnl      option) any later version.
+dnl
+dnl  or
+dnl
+dnl    * the GNU General Public License as published by the Free Software
+dnl      Foundation; either version 2 of the License, or (at your option) any
+dnl      later version.
+dnl
+dnl  or both in parallel, as here.
+dnl
+dnl  The GNU MP Library is distributed in the hope that it will be useful, but
+dnl  WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
+dnl  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
+dnl  for more details.
+dnl
+dnl  You should have received copies of the GNU General Public License and the
+dnl  GNU Lesser General Public License along with the GNU MP Library.  If not,
+dnl  see https://www.gnu.org/licenses/.
+
+include(`../config.m4')
+
+
+C ev67: 2.5 cycles/limb
+
+
+C unsigned long mpn_hamdist (mp_srcptr xp, mp_srcptr yp, mp_size_t size);
+C
+C The hope was for 2.0 c/l here, but that isn't achieved.  We're limited by
+C renaming register shortage.  Since we need 5 instructions per limb, further
+C unrolling could approach 1.5 c/l.
+C
+C The main loop processes two limbs from each operand on each iteration.  An
+C odd size is handled by processing xp[0]^yp[0] at the start.  If the size
+C is even that result is discarded, and is repeated by the main loop.
+C
+
+ASM_START()
+PROLOGUE(mpn_hamdist)
+
+	C r16	xp
+	C r17	yp
+	C r18	size
+
+	ldq	r1, 0(r16)		C L0  xp[0]
+	ldq	r2, 0(r17)		C L1  yp[0]
+	and	r18, 1, r8		C U1  1 if size odd
+	srl	r18, 1, r18		C U0  size, limb pairs
+
+	clr	r0			C L0  initial total
+	s8addq	r8, r17, r17		C U1  yp++ if size odd
+	s8addq	r8, r16, r16		C L1  xp++ if size odd
+	clr	r6			C U0  dummy initial xor 1
+
+	xor	r1, r2, r5		C L   initial xor 0
+	beq	r18, L(one)		C U   if size==1
+
+	cmoveq	r8, r31, r5		C L   discard first limb if size even
+	unop				C U
+
+
+	ALIGN(16)
+L(top):
+	C r0	total accumulating
+	C r7	xor 0
+	C r8	xor 1
+	C r16	xp, incrementing
+	C r17	yp, incrementing
+	C r18	size, limb pairs, decrementing
+
+	ldq	r1, 0(r16)		C L
+	ldq	r2, 0(r17)		C L
+	ctpop	r5, r7			C U0
+	lda	r16, 16(r16)		C U
+
+	ldq	r3, -8(r16)		C L
+	ldq	r4, 8(r17)		C L
+	ctpop	r6, r8			C U0
+	lda	r17, 16(r17)		C U
+
+	ldl	r31, 256(r16)		C L	prefetch
+	ldl	r31, 256(r17)		C L	prefetch
+	xor	r1, r2, r5		C U
+	lda	r18, -1(r18)		C U
+
+	xor	r3, r4, r6		C U
+	addq	r0, r7, r0		C L
+	addq	r0, r8, r0		C L
+	bne	r18, L(top)		C U
+
+
+	ctpop	r6, r8			C U0
+	addq	r0, r8, r0		C L
+L(one):
+	ctpop	r5, r7			C U0
+	addq	r0, r7, r0		C L
+
+	ret	r31, (r26), 1		C L0
+
+EPILOGUE()
+ASM_END()
diff --git a/vendor/gmp-6.3.0/mpn/alpha/ev67/popcount.asm b/vendor/gmp-6.3.0/mpn/alpha/ev67/popcount.asm
new file mode 100644
index 0000000..049c1cd
--- /dev/null
+++ b/vendor/gmp-6.3.0/mpn/alpha/ev67/popcount.asm
@@ -0,0 +1,101 @@
+dnl  Alpha ev67 mpn_popcount -- mpn bit population count.
+
+dnl  Copyright 2003, 2005 Free Software Foundation, Inc.
+
+dnl  This file is part of the GNU MP Library.
+dnl
+dnl  The GNU MP Library is free software; you can redistribute it and/or modify
+dnl  it under the terms of either:
+dnl
+dnl    * the GNU Lesser General Public License as published by the Free
+dnl      Software Foundation; either version 3 of the License, or (at your
+dnl      option) any later version.
+dnl
+dnl  or
+dnl
+dnl    * the GNU General Public License as published by the Free Software
+dnl      Foundation; either version 2 of the License, or (at your option) any
+dnl      later version.
+dnl
+dnl  or both in parallel, as here.
+dnl
+dnl  The GNU MP Library is distributed in the hope that it will be useful, but
+dnl  WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
+dnl  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
+dnl  for more details.
+dnl
+dnl  You should have received copies of the GNU General Public License and the
+dnl  GNU Lesser General Public License along with the GNU MP Library.  If not,
+dnl  see https://www.gnu.org/licenses/.
+
+include(`../config.m4')
+
+
+C ev67: 1.5 cycles/limb
+
+
+C unsigned long mpn_popcount (mp_srcptr src, mp_size_t size);
+C
+C This schedule seems necessary for the full 1.5 c/l, the IQ can't quite hide
+C all latencies, the addq's must be deferred to the next iteration.
+C
+C Since we need just 3 instructions per limb, further unrolling could approach
+C 1.0 c/l.
+C
+C The main loop processes two limbs at a time.  An odd size is handled by
+C processing src[0] at the start.  If the size is even that result is
+C discarded, and src[0] is repeated by the main loop.
+C
+
+ASM_START()
+PROLOGUE(mpn_popcount)
+
+	C r16	src
+	C r17	size
+
+	ldq	r0, 0(r16)		C L0  src[0]
+	and	r17, 1, r8		C U1  1 if size odd
+	srl	r17, 1, r17		C U0  size, limb pairs
+
+	s8addq	r8, r16, r16		C L1  src++ if size odd
+	ctpop	r0, r0			C U0
+	beq	r17, L(one)		C U1  if size==1
+
+	cmoveq	r8, r31, r0		C L   discard first limb if size even
+	clr	r3			C L
+
+	clr	r4			C L
+	unop				C U
+	unop				C L
+	unop				C U
+
+
+	ALIGN(16)
+L(top):
+	C r0	total accumulating
+	C r3	pop 0
+	C r4	pop 1
+	C r16	src, incrementing
+	C r17	size, decrementing
+
+	ldq	r1, 0(r16)		C L
+	ldq	r2, 8(r16)		C L
+	lda	r16, 16(r16)		C U
+	lda	r17, -1(r17)		C U
+
+	addq	r0, r3, r0		C L
+	addq	r0, r4, r0		C L
+	ctpop	r1, r3			C U0
+	ctpop	r2, r4			C U0
+
+	ldl	r31, 512(r16)		C L	prefetch
+	bne	r17, L(top)		C U
+
+
+	addq	r0, r3, r0		C L
+	addq	r0, r4, r0		C U
+L(one):
+	ret	r31, (r26), 1		C L0
+
+EPILOGUE()
+ASM_END()