diff options
Diffstat (limited to 'gmp/mpn/alpha/ev67')
| -rw-r--r-- | gmp/mpn/alpha/ev67/gcd_1.asm | 145 | ||||
| -rw-r--r-- | gmp/mpn/alpha/ev67/hamdist.asm | 111 | ||||
| -rw-r--r-- | gmp/mpn/alpha/ev67/popcount.asm | 101 |
3 files changed, 357 insertions, 0 deletions
diff --git a/gmp/mpn/alpha/ev67/gcd_1.asm b/gmp/mpn/alpha/ev67/gcd_1.asm new file mode 100644 index 0000000000..55fa7d3673 --- /dev/null +++ b/gmp/mpn/alpha/ev67/gcd_1.asm @@ -0,0 +1,145 @@ +dnl Alpha ev67 mpn_gcd_1 -- 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. +C +C Enhancement: +C +C On the jsr, !lituse_jsr! (when available) would allow the linker to relax +C it to a bsr, but probably only in a static binary. Plain "jsr foo" gives +C the right object code for relaxation, and ought to be available +C everywhere, but we prefer to schedule the GOT ldq (LEA) back earlier, for +C the usual case of running in a shared library. +C +C bsr could perhaps be used explicitly anyway. We should be able to assume +C modexact is in the same module as us (ie. shared library or mainline). +C Would there be any worries about the size of the displacement? Could +C always put modexact and gcd_1 in the same .o to be certain. + +ASM_START() +PROLOGUE(mpn_gcd_1, gp) + + C r16 xp + C r17 size + C r18 y + + C ldah C l + C lda C u + + ldq r0, 0(r16) C L x = xp[0] + lda r30, -32(r30) C u alloc stack + + LEA( r27, mpn_modexact_1c_odd) C L modexact addr, ldq (gp) + stq r10, 16(r30) C L save r10 + cttz r18, r10 C U0 y twos + cmpeq r17, 1, r5 C u test size==1 + + stq r9, 8(r30) C L save r9 + clr r19 C u zero c for modexact + unop + unop + + cttz r0, r6 C U0 x twos + stq r26, 0(r30) C L save ra + + srl r18, r10, r18 C U y odd + + mov r18, r9 C l hold y across call + + cmpult r6, r10, r2 C u test x_twos < y_twos + + cmovne r2, r6, r10 C l common_twos = min(x_twos,y_twos) + bne r5, L(one) C U no modexact if size==1 + jsr r26, (r27), mpn_modexact_1c_odd C L0 + + LDGP( r29, 0(r26)) C u,l ldah,lda + cttz r0, r6 C U0 new x twos + ldq r26, 0(r30) C L restore ra + +L(one): + mov r9, r1 C u y + ldq r9, 8(r30) C L restore r9 + mov r10, r2 C u common twos + ldq r10, 16(r30) C L restore r10 + + lda r30, 32(r30) C l free stack + beq r0, L(done) C U return y if x%y==0 + + srl r0, r6, r0 C U x odd + unop + + ALIGN(16) +L(top): + C r0 x + C r1 y + C r2 common twos, for use at end + + 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(done): + sll r1, r2, r0 C U0 return y << common_twos + ret r31, (r26), 1 C L0 + +EPILOGUE() +ASM_END() diff --git a/gmp/mpn/alpha/ev67/hamdist.asm b/gmp/mpn/alpha/ev67/hamdist.asm new file mode 100644 index 0000000000..4b13e9f14f --- /dev/null +++ b/gmp/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/gmp/mpn/alpha/ev67/popcount.asm b/gmp/mpn/alpha/ev67/popcount.asm new file mode 100644 index 0000000000..049c1cd239 --- /dev/null +++ b/gmp/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() |