head 1.2; access; symbols perseant-exfatfs-base-20250801:1.2 perseant-exfatfs-base-20240630:1.2 perseant-exfatfs:1.2.0.2 perseant-exfatfs-base:1.2 netbsd-8-3-RELEASE:1.1.1.2 netbsd-9-4-RELEASE:1.1.1.3 netbsd-9-3-RELEASE:1.1.1.3 cjep_sun2x-base1:1.1.1.3 cjep_sun2x:1.1.1.3.0.12 cjep_sun2x-base:1.1.1.3 cjep_staticlib_x-base1:1.1.1.3 netbsd-9-2-RELEASE:1.1.1.3 cjep_staticlib_x:1.1.1.3.0.10 cjep_staticlib_x-base:1.1.1.3 netbsd-9-1-RELEASE:1.1.1.3 gmp-6-2-0:1.1.1.3 phil-wifi-20200421:1.1.1.3 phil-wifi-20200411:1.1.1.3 is-mlppp:1.1.1.3.0.8 is-mlppp-base:1.1.1.3 phil-wifi-20200406:1.1.1.3 netbsd-8-2-RELEASE:1.1.1.2 netbsd-9-0-RELEASE:1.1.1.3 netbsd-9-0-RC2:1.1.1.3 netbsd-9-0-RC1:1.1.1.3 phil-wifi-20191119:1.1.1.3 netbsd-9:1.1.1.3.0.6 netbsd-9-base:1.1.1.3 phil-wifi-20190609:1.1.1.3 netbsd-8-1-RELEASE:1.1.1.2 netbsd-8-1-RC1:1.1.1.2 pgoyette-compat-merge-20190127:1.1.1.3 pgoyette-compat-20190127:1.1.1.3 pgoyette-compat-20190118:1.1.1.3 pgoyette-compat-1226:1.1.1.3 pgoyette-compat-1126:1.1.1.3 pgoyette-compat-1020:1.1.1.3 pgoyette-compat-0930:1.1.1.3 pgoyette-compat-0906:1.1.1.3 netbsd-7-2-RELEASE:1.1.1.2 pgoyette-compat-0728:1.1.1.3 netbsd-8-0-RELEASE:1.1.1.2 phil-wifi:1.1.1.3.0.4 phil-wifi-base:1.1.1.3 pgoyette-compat-0625:1.1.1.3 netbsd-8-0-RC2:1.1.1.2 pgoyette-compat-0521:1.1.1.3 pgoyette-compat-0502:1.1.1.3 pgoyette-compat-0422:1.1.1.3 netbsd-8-0-RC1:1.1.1.2 pgoyette-compat-0415:1.1.1.3 pgoyette-compat-0407:1.1.1.3 pgoyette-compat-0330:1.1.1.3 pgoyette-compat-0322:1.1.1.3 pgoyette-compat-0315:1.1.1.3 netbsd-7-1-2-RELEASE:1.1.1.2 pgoyette-compat:1.1.1.3.0.2 pgoyette-compat-base:1.1.1.3 netbsd-7-1-1-RELEASE:1.1.1.2 matt-nb8-mediatek:1.1.1.2.0.22 matt-nb8-mediatek-base:1.1.1.2 gmp-6-1-2:1.1.1.3 perseant-stdc-iso10646:1.1.1.2.0.20 perseant-stdc-iso10646-base:1.1.1.2 netbsd-8:1.1.1.2.0.18 netbsd-8-base:1.1.1.2 prg-localcount2-base3:1.1.1.2 prg-localcount2-base2:1.1.1.2 prg-localcount2-base1:1.1.1.2 prg-localcount2:1.1.1.2.0.16 prg-localcount2-base:1.1.1.2 pgoyette-localcount-20170426:1.1.1.2 bouyer-socketcan-base1:1.1.1.2 pgoyette-localcount-20170320:1.1.1.2 netbsd-7-1:1.1.1.2.0.14 netbsd-7-1-RELEASE:1.1.1.2 netbsd-7-1-RC2:1.1.1.2 netbsd-7-nhusb-base-20170116:1.1.1.2 bouyer-socketcan:1.1.1.2.0.12 bouyer-socketcan-base:1.1.1.2 pgoyette-localcount-20170107:1.1.1.2 netbsd-7-1-RC1:1.1.1.2 pgoyette-localcount-20161104:1.1.1.2 netbsd-7-0-2-RELEASE:1.1.1.2 localcount-20160914:1.1.1.2 netbsd-7-nhusb:1.1.1.2.0.10 netbsd-7-nhusb-base:1.1.1.2 pgoyette-localcount-20160806:1.1.1.2 pgoyette-localcount-20160726:1.1.1.2 pgoyette-localcount:1.1.1.2.0.8 pgoyette-localcount-base:1.1.1.2 netbsd-7-0-1-RELEASE:1.1.1.2 netbsd-7-0:1.1.1.2.0.6 netbsd-7-0-RELEASE:1.1.1.2 netbsd-7-0-RC3:1.1.1.2 netbsd-7-0-RC2:1.1.1.2 netbsd-7-0-RC1:1.1.1.2 netbsd-6-0-6-RELEASE:1.1.1.1 netbsd-6-1-5-RELEASE:1.1.1.1 netbsd-7:1.1.1.2.0.4 netbsd-7-base:1.1.1.2 yamt-pagecache-base9:1.1.1.2 yamt-pagecache-tag8:1.1.1.1 netbsd-6-1-4-RELEASE:1.1.1.1 netbsd-6-0-5-RELEASE:1.1.1.1 tls-earlyentropy:1.1.1.2.0.2 tls-earlyentropy-base:1.1.1.2 riastradh-xf86-video-intel-2-7-1-pre-2-21-15:1.1.1.2 riastradh-drm2-base3:1.1.1.2 netbsd-6-1-3-RELEASE:1.1.1.1 netbsd-6-0-4-RELEASE:1.1.1.1 gmp-5-1-3:1.1.1.2 netbsd-6-1-2-RELEASE:1.1.1.1 netbsd-6-0-3-RELEASE:1.1.1.1 netbsd-6-1-1-RELEASE:1.1.1.1 riastradh-drm2-base2:1.1.1.1 riastradh-drm2-base1:1.1.1.1 riastradh-drm2:1.1.1.1.0.12 riastradh-drm2-base:1.1.1.1 netbsd-6-1:1.1.1.1.0.16 netbsd-6-0-2-RELEASE:1.1.1.1 netbsd-6-1-RELEASE:1.1.1.1 netbsd-6-1-RC4:1.1.1.1 netbsd-6-1-RC3:1.1.1.1 agc-symver:1.1.1.1.0.14 agc-symver-base:1.1.1.1 netbsd-6-1-RC2:1.1.1.1 netbsd-6-1-RC1:1.1.1.1 yamt-pagecache-base8:1.1.1.1 netbsd-6-0-1-RELEASE:1.1.1.1 yamt-pagecache-base7:1.1.1.1 matt-nb6-plus-nbase:1.1.1.1 yamt-pagecache-base6:1.1.1.1 netbsd-6-0:1.1.1.1.0.10 netbsd-6-0-RELEASE:1.1.1.1 netbsd-6-0-RC2:1.1.1.1 tls-maxphys:1.1.1.1.0.8 tls-maxphys-base:1.1.1.2 matt-nb6-plus:1.1.1.1.0.6 matt-nb6-plus-base:1.1.1.1 netbsd-6-0-RC1:1.1.1.1 yamt-pagecache-base5:1.1.1.1 yamt-pagecache-base4:1.1.1.1 netbsd-6:1.1.1.1.0.4 netbsd-6-base:1.1.1.1 yamt-pagecache-base3:1.1.1.1 yamt-pagecache-base2:1.1.1.1 yamt-pagecache:1.1.1.1.0.2 yamt-pagecache-base:1.1.1.1 gmp-5-0-2:1.1.1.1 gmp:1.1.1; locks; strict; comment @;; @; 1.2 date 2021.07.11.21.15.46; author mrg; state dead; branches; next 1.1; commitid UkDd4YZ64jVf9C0D; 1.1 date 2011.06.20.05.54.39; author mrg; state Exp; branches 1.1.1.1; next ; 1.1.1.1 date 2011.06.20.05.54.39; author mrg; state Exp; branches 1.1.1.1.2.1 1.1.1.1.8.1; next 1.1.1.2; 1.1.1.2 date 2013.11.29.07.49.48; author mrg; state Exp; branches; next 1.1.1.3; commitid L2Av4PuGmdoL39fx; 1.1.1.3 date 2017.08.22.09.40.49; author mrg; state Exp; branches; next ; commitid W5kmAIk8hwVpSb4A; 1.1.1.1.2.1 date 2014.05.22.14.09.01; author yamt; state Exp; branches; next ; commitid nx2BSsHy0NPeAxBx; 1.1.1.1.8.1 date 2014.08.19.23.59.50; author tls; state Exp; branches; next ; commitid jTnpym9Qu0o4R1Nx; desc @@ 1.2 log @merge GMP 6.2.1. @ text @dnl Itanium-2 mpn_gcd_1 -- mpn by 1 gcd. dnl Copyright 2002, 2003, 2004, 2005 Free Software Foundation, Inc. dnl This file is part of the GNU MP Library. dnl The GNU MP Library is free software; you can redistribute it and/or modify dnl it under the terms of the GNU Lesser General Public License as published dnl by the Free Software Foundation; either version 3 of the License, or (at dnl your option) any later version. 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 Lesser General Public dnl License for more details. dnl You should have received a copy of the GNU Lesser General Public License dnl along with the GNU MP Library. If not, see http://www.gnu.org/licenses/. include(`../config.m4') C cycles/bitpair (1x1 gcd) C Itanium: 14 (approx) C Itanium 2: 6.3 C mpn_gcd_1 (mp_srcptr xp, mp_size_t xsize, mp_limb_t y); C C The entry sequence is designed to expect xsize>1 and hence a modexact C call. This ought to be more common than a 1x1 operation. Our critical C path is thus stripping factors of 2 from y, calling modexact, then C stripping factors of 2 from the x remainder returned. C C The common factors of 2 between x and y must be determined using the C original x, not the remainder from the modexact. This is done with C x_orig which is xp[0]. There's plenty of time to do this while the rest C of the modexact etc is happening. C C It's possible xp[0] is zero. In this case the trailing zeros calculation C popc((x-1)&~x) gives 63, and that's clearly no less than what y will C have, making min(x_twos,y_twos) == y_twos. C C The main loop consists of transforming x,y to abs(x-y),min(x,y), and then C stripping factors of 2 from abs(x-y). Those factors of two are C determined from just y-x, without the abs(), since there's the same C number of trailing zeros on n or -n in twos complement. That makes the C dependent chain C C cycles C 1 sub x-y and x-y-1 C 3 andcm (x-y-1)&~(x-y) C 2 popcnt trailing zeros C 3 shr.u strip abs(x-y) C --- C 9 C C The selection of x-y versus y-x for abs(x-y), and the selection of the C minimum of x and y, is done in parallel with the above. C C The algorithm takes about 0.68 iterations per bit (two N bit operands) on C average, hence the final 6.3 cycles/bitpair. C C The loop is not as fast as one might hope, since there's extra latency C from andcm going across to the `multimedia' popcnt, and vice versa from C multimedia shr.u back to the integer sub. C C The loop branch is .sptk.clr since we usually expect a good number of C iterations, and the iterations are data dependent so it's unlikely past C results will predict anything much about the future. C C Not done: C C An alternate algorithm which didn't strip all twos, but instead applied C tbit and predicated extr on x, and then y, was attempted. The loop was 6 C cycles, but the algorithm is an average 1.25 iterations per bitpair for a C total 7.25 c/bp, which is slower than the current approach. C C Alternatives: C C Perhaps we could do something tricky by extracting a few high bits and a C few low bits from the operands, and looking up a table which would give a C set of predicates to control some shifts or subtracts or whatever. That C could knock off multiple bits per iteration. C C The right shifts are a bit of a bottleneck (shr at 2 or 3 cycles, or extr C only going down I0), perhaps it'd be possible to shift left instead, C using add. That would mean keeping track of the lowest not-yet-zeroed C bit, using some sort of mask. C C Itanium-1: C C This code is not designed for itanium-1 and in fact doesn't run well on C that chip. The loop seems to be about 21 cycles, probably because we end C up with a 10 cycle replay for not forcibly scheduling the shr.u latency. C Lack of branch hints might introduce a couple of bubbles too. C ASM_START() .explicit C What does this mean? C HP's assembler requires these declarations for importing mpn_modexact_1c_odd .global mpn_modexact_1c_odd .type mpn_modexact_1c_odd,@@function PROLOGUE(mpn_gcd_1) C r32 xp C r33 xsize C r34 y define(x, r8) define(xp_orig, r32) define(xsize, r33) define(y, r34) define(inputs, 3) define(save_rp, r35) define(save_pfs, r36) define(x_orig, r37) define(x_orig_one, r38) define(y_twos, r39) define(locals, 5) define(out_xp, r40) define(out_xsize, r41) define(out_divisor, r42) define(out_carry, r43) define(outputs, 4) .prologue { .mmi; ifdef(`HAVE_ABI_32', ` addp4 r9 = 0, xp_orig define(xp,r9)', C M0 ` define(xp,xp_orig)') .save ar.pfs, save_pfs alloc save_pfs = ar.pfs, inputs, locals, outputs, 0 C M2 .save rp, save_rp mov save_rp = b0 C I0 }{ .body add r10 = -1, y C M3 y-1 } ;; { .mmi; ld8 x = [xp] C M0 x = xp[0] if no modexact ld8 x_orig = [xp] C M1 orig x for common twos cmp.ne p6,p0 = 1, xsize C I0 }{ .mmi; andcm y_twos = r10, y C M2 (y-1)&~y mov out_xp = xp_orig C M3 mov out_xsize = xsize C I1 } ;; mov out_carry = 0 C popcnt y_twos = y_twos C I0 y twos ;; C { .mmi; add x_orig_one = -1, x_orig C M0 orig x-1 shr.u out_divisor = y, y_twos C I0 y without twos }{ shr.u y = y, y_twos C I1 y without twos (p6) br.call.sptk.many b0 = mpn_modexact_1c_odd C if xsize>1 } ;; C modexact can leave x==0 { .mmi; cmp.eq p6,p0 = 0, x C M0 if {xp,xsize} % y == 0 andcm x_orig = x_orig_one, x_orig C M1 orig (x-1)&~x add r9 = -1, x C I0 x-1 } ;; { .mmi; andcm r9 = r9, x C M0 (x-1)&~x mov b0 = save_rp C I0 } ;; C popcnt x_orig = x_orig C I0 orig x twos popcnt r9 = r9 C I0 x twos ;; C { cmp.lt p7,p0 = x_orig, y_twos C M0 orig x_twos < y_twos shr.u x = x, r9 C I0 x odd } ;; { (p7) mov y_twos = x_orig C M0 common twos add r10 = -1, y C I0 y-1 (p6) br.dpnt.few .Ldone_y C B0 x%y==0 then result y } ;; C C No noticable difference in speed for the loop aligned to C 32 or just 16. .Ltop: C r8 x C r10 y-1 C r34 y C r38 common twos, for use at end { .mmi; cmp.gtu p8,p9 = x, y C M0 x>y cmp.ne p10,p0 = x, y C M1 x==y sub r9 = y, x C I0 d = y - x }{ .mmi; sub r10 = r10, x C M2 d-1 = y - x - 1 } ;; { .mmi; .pred.rel "mutex", p8, p9 (p8) sub x = x, y C M0 x>y use x=x-y, y unchanged (p9) mov y = x C M1 y>=x use y=x (p9) mov x = r9 C I0 y>=x use x=y-x }{ .mmi; andcm r9 = r10, r9 C M2 (d-1)&~d ;; add r10 = -1, y C M0 new y-1 popcnt r9 = r9 C I0 twos on x-y } ;; { shr.u x = x, r9 C I0 new x without twos (p10) br.sptk.few.clr .Ltop } ;; C result is y .Ldone_y: shl r8 = y, y_twos C I common factors of 2 ;; mov ar.pfs = save_pfs C I0 br.ret.sptk.many b0 EPILOGUE() @ 1.1 log @Initial revision @ text @@ 1.1.1.1 log @initial import of GMP 5.0.2. GNU MP is a library for arbitrary precision arithmetic, operating on signed integers, rational numbers, and floating point numbers. It has a rich set of functions, and the functions have a regular interface. GMP is necessary for GCC >= 4.2. @ text @@ 1.1.1.1.8.1 log @Rebase to HEAD as of a few days ago. @ text @d3 1 a3 4 dnl Contributed to the GNU project by Kevin Ryde, innerloop by Torbjorn dnl Granlund. dnl Copyright 2002, 2003, 2004, 2005, 2012 Free Software Foundation, Inc. d24 2 a25 2 C Itanium: ? C Itanium 2: 5.8 (trimmable to 5.64 with huge ctz_table) d48 9 a56 1 C dependent chain 8 cycles deep. d59 1 a59 1 C minimum of x and y, is done in parallel with the critical path. d62 9 a70 1 C average, hence the final 5.8 cycles/bitpair. d91 7 a97 4 C TODO: C * Once mod_1_N exists in assembly for Itanium, add conditional calls. C * Call bmod_1 even for n=1 when up[0] >> v0 (like other gcd_1 impls). C * Probably avoid popcnt also outside of loop, instead use ctz_table. a105 12 C ctz_table[n] is the number of trailing zeros on n, or MAXSHIFT if n==0. deflit(MAXSHIFT, 7) deflit(MASK, eval((m4_lshift(1,MAXSHIFT))-1)) .section ".rodata" ctz_table: .byte MAXSHIFT forloop(i,1,MASK, ` .byte m4_count_trailing_zeros(i) ') d149 2 d154 2 d172 2 d175 1 d179 2 d187 1 a187 1 (p6) br.dpnt.few L(done_y) C B0 x%y==0 then result y d190 22 a211 1 addl r22 = @@ltoffx(ctz_table#), r1 a212 2 ld8.mov r22 = [r22], ctz_table# br L(ent) d214 7 a221 26 ALIGN(32) L(top): .pred.rel "mutex", p6,p7 .mmi; and r20 = MASK, r19 (p7) mov y = x (p6) sub x = x, y .mmi; (p7) mov x = r19 nop 0 nop 0 ;; L(mid): .mmb; add r21 = r22, r20 cmp.eq p10,p0 = 0, r20 (p10) br.spnt.few.clr L(shift_alot) ;; .mmi; ld1 r16 = [r21] ;; nop 0 shr.u x = x, r16 ;; L(ent): .mmi; sub r19 = y, x cmp.gtu p6,p7 = x, y cmp.ne p8,p0 = x, y .mmb; nop 0 nop 0 (p8) br.sptk.few.clr L(top) d225 3 a227 1 L(done_y): a228 1 shl r8 = y, y_twos C I common factors of 2 a230 4 L(shift_alot): extr.u r20 = x, MAXSHIFT, MAXSHIFT shr.u x = x, MAXSHIFT br L(mid) @ 1.1.1.1.2.1 log @sync with head. for a reference, the tree before this commit was tagged as yamt-pagecache-tag8. this commit was splitted into small chunks to avoid a limitation of cvs. ("Protocol error: too many arguments") @ text @d3 1 a3 4 dnl Contributed to the GNU project by Kevin Ryde, innerloop by Torbjorn dnl Granlund. dnl Copyright 2002, 2003, 2004, 2005, 2012 Free Software Foundation, Inc. d24 2 a25 2 C Itanium: ? C Itanium 2: 5.8 (trimmable to 5.64 with huge ctz_table) d48 9 a56 1 C dependent chain 8 cycles deep. d59 1 a59 1 C minimum of x and y, is done in parallel with the critical path. d62 9 a70 1 C average, hence the final 5.8 cycles/bitpair. d91 7 a97 4 C TODO: C * Once mod_1_N exists in assembly for Itanium, add conditional calls. C * Call bmod_1 even for n=1 when up[0] >> v0 (like other gcd_1 impls). C * Probably avoid popcnt also outside of loop, instead use ctz_table. a105 12 C ctz_table[n] is the number of trailing zeros on n, or MAXSHIFT if n==0. deflit(MAXSHIFT, 7) deflit(MASK, eval((m4_lshift(1,MAXSHIFT))-1)) .section ".rodata" ctz_table: .byte MAXSHIFT forloop(i,1,MASK, ` .byte m4_count_trailing_zeros(i) ') d149 2 d154 2 d172 2 d175 1 d179 2 d187 1 a187 1 (p6) br.dpnt.few L(done_y) C B0 x%y==0 then result y d190 22 a211 1 addl r22 = @@ltoffx(ctz_table#), r1 a212 2 ld8.mov r22 = [r22], ctz_table# br L(ent) d214 7 a221 26 ALIGN(32) L(top): .pred.rel "mutex", p6,p7 .mmi; and r20 = MASK, r19 (p7) mov y = x (p6) sub x = x, y .mmi; (p7) mov x = r19 nop 0 nop 0 ;; L(mid): .mmb; add r21 = r22, r20 cmp.eq p10,p0 = 0, r20 (p10) br.spnt.few.clr L(shift_alot) ;; .mmi; ld1 r16 = [r21] ;; nop 0 shr.u x = x, r16 ;; L(ent): .mmi; sub r19 = y, x cmp.gtu p6,p7 = x, y cmp.ne p8,p0 = x, y .mmb; nop 0 nop 0 (p8) br.sptk.few.clr L(top) d225 3 a227 1 L(done_y): a228 1 shl r8 = y, y_twos C I common factors of 2 a230 4 L(shift_alot): extr.u r20 = x, MAXSHIFT, MAXSHIFT shr.u x = x, MAXSHIFT br L(mid) @ 1.1.1.2 log @initial import GMP 5.1.3 sources. changes include: fixes for: - mpn_sbpi1_div_qr_sec and mpn_sbpi1_div_r_sec - mpz_powm_ui - AMD family 11h - mpz_powm_sec and mpn_powm_sec - ASSERT() fixes - gcd, gcdext, and invert function fixes - some PPC division operations @ text @d3 1 a3 4 dnl Contributed to the GNU project by Kevin Ryde, innerloop by Torbjorn dnl Granlund. dnl Copyright 2002, 2003, 2004, 2005, 2012 Free Software Foundation, Inc. d24 2 a25 2 C Itanium: ? C Itanium 2: 5.8 (trimmable to 5.64 with huge ctz_table) d48 9 a56 1 C dependent chain 8 cycles deep. d59 1 a59 1 C minimum of x and y, is done in parallel with the critical path. d62 9 a70 1 C average, hence the final 5.8 cycles/bitpair. d91 7 a97 4 C TODO: C * Once mod_1_N exists in assembly for Itanium, add conditional calls. C * Call bmod_1 even for n=1 when up[0] >> v0 (like other gcd_1 impls). C * Probably avoid popcnt also outside of loop, instead use ctz_table. a105 12 C ctz_table[n] is the number of trailing zeros on n, or MAXSHIFT if n==0. deflit(MAXSHIFT, 7) deflit(MASK, eval((m4_lshift(1,MAXSHIFT))-1)) .section ".rodata" ctz_table: .byte MAXSHIFT forloop(i,1,MASK, ` .byte m4_count_trailing_zeros(i) ') d149 2 d154 2 d172 2 d175 1 d179 2 d187 1 a187 1 (p6) br.dpnt.few L(done_y) C B0 x%y==0 then result y d190 22 a211 1 addl r22 = @@ltoffx(ctz_table#), r1 a212 2 ld8.mov r22 = [r22], ctz_table# br L(ent) d214 7 a221 26 ALIGN(32) L(top): .pred.rel "mutex", p6,p7 .mmi; and r20 = MASK, r19 (p7) mov y = x (p6) sub x = x, y .mmi; (p7) mov x = r19 nop 0 nop 0 ;; L(mid): .mmb; add r21 = r22, r20 cmp.eq p10,p0 = 0, r20 (p10) br.spnt.few.clr L(shift_alot) ;; .mmi; ld1 r16 = [r21] ;; nop 0 shr.u x = x, r16 ;; L(ent): .mmi; sub r19 = y, x cmp.gtu p6,p7 = x, y cmp.ne p8,p0 = x, y .mmb; nop 0 nop 0 (p8) br.sptk.few.clr L(top) d225 3 a227 1 L(done_y): a228 1 shl r8 = y, y_twos C I common factors of 2 a230 4 L(shift_alot): extr.u r20 = x, MAXSHIFT, MAXSHIFT shr.u x = x, MAXSHIFT br L(mid) @ 1.1.1.3 log @initial import of GMP 6.1.2. main changes from 5.1.3 below. notes: - support for thumb-less ARM chips was in our port of 5.1.3, but a similar method has been provided upstream now - someone should look at the AVX failure reports, and fix them Changes between GMP version 6.1.0 and 6.1.1 FEATURES * Work around faulty cpuid on some recent Intel chips (this allows GMP to run on Skylake Pentiums). * Support thumb-less ARM chips. Changes between GMP version 6.0.* and 6.1.0 BUGS FIXED * The public function mpn_com is now correctly declared in gmp.h. * Healed possible failures of mpn_sec_sqr for non-cryptographic sizes for some obsolete CPUs. * Various problems related to precision for mpf have been fixed. * Fixed ABI incompatible stack alignment in calls from assembly code. * Fixed PIC bug in popcount affecting Intel processors using the 32-bit ABI. SPEEDUPS * Speedup for Intel Broadwell and Skylake through assembly code making use of new ADX instructions. * Square root is now faster when the remainder is not needed. Also the speed to compute the k-th root improved, for small sizes. FEATURES * New C++ functions gcd and lcm for mpz_class. * New public mpn functions mpn_divexact_1, mpn_zero_p, and mpn_cnd_swap. * New public mpq_cmp_z function, to efficiently compare rationals with integers. * Support for more 32-bit arm processors. * Support for AVX-less modern x86 CPUs. (Such support might be missing either because the CPU vendor chose to disable AVX, or because the running kernel lacks AVX context switch support.) * Support for NetBSD under Xen; we switch off AVX unconditionally under NetBSD since a bug in NetBSD makes AVX fail under Xen. MISC * Tuned values for FFT multiplications are provided for larger number on many platforms. Changes between GMP version 5.1.* and 6.0.0 BUGS FIXED * The function mpz_invert now considers any number invertible in Z/1Z. * The mpn multiply code now handles operands of more than 2^31 limbs correctly. (Note however that the mpz code is limited to 2^32 bits on 32-bit hosts and 2^37 bits on 64-bit hosts.) SPEEDUPS * Plain division of large operands is faster and more monotonous in operand size. * Major speedup for ARM, in particular ARM Cortex-A15, thanks to improved assembly. * Speedup for Intel Sandy Bridge, Ivy Bridge, Haswell, thanks to rewritten and vastly expanded assembly support. Speedup also for the older Core 2 and Nehalem. * Faster mixed arithmetic between mpq_class and double. FEATURES * Support for new Intel and AMD CPUs. * New public functions mpn_sec_mul and mpn_sec_sqr, implementing side-channel silent multiplication and squaring. * New public functions mpn_sec_div_qr and mpn_sec_div_r, implementing side-channel silent division. * New public functions mpn_cnd_add_n and mpn_cnd_sub_n. Side-channel silent conditional addition and subtraction. * New public function mpn_sec_powm, implementing side-channel silent modexp. * New public function mpn_sec_invert, implementing side-channel silent modular inversion. * Better support for applications which use the mpz_t type, but nevertheless need to call some of the lower-level mpn functions. See the documentation for mpz_limbs_read and related functions. @ text @d6 1 a6 1 dnl Copyright 2002-2005, 2012, 2013, 2015 Free Software Foundation, Inc. d9 1 a9 1 dnl d11 4 a14 14 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 d17 5 a21 6 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/. d28 1 a28 1 C Itanium 2: 5.1 d95 1 a95 3 C .section ".rodata" .rodata ALIGN(m4_lshift(1,MAXSHIFT)) C align table to allow using dep d97 1 a97 1 data1 MAXSHIFT d99 1 a99 1 ` data1 m4_count_trailing_zeros(i) d123 1 a123 1 {.mmi; d130 21 a150 23 mov save_rp = b0 C I0 }{.mbb; .body add r10 = -1, y C M3 y-1 nop.b 0 C B0 nop.b 0 C B1 ;; }{.mmi; ld8 x = [xp] C M0 x = xp[0] if no modexact ld8 x_orig = [xp] C M1 orig x for common twos cmp.ne p6,p0 = 1, xsize C I0 }{.mmi; andcm y_twos = r10, y C M2 (y-1)&~y mov out_xp = xp_orig C M3 mov out_xsize = xsize C I1 ;; }{.mmi; mov out_carry = 0 C M0 nop.m 0 C M1 popcnt y_twos = y_twos C I0 y twos ;; }{.mmi; add x_orig_one = -1, x_orig C M0 orig x-1 nop.m 0 C M1 shr.u out_divisor = y, y_twos C I0 y without twos }{.mib; nop.m 0 C M2 shr.u y = y, y_twos C I1 y without twos d152 4 a155 4 ;; } C modexact can leave x==0 {.mmi; cmp.eq p6,p0 = 0, x C M0 if {xp,xsize} % y == 0 d157 16 a172 16 add r9 = -1, x C I0 x-1 ;; }{.mmi; andcm r9 = r9, x C M0 (x-1)&~x nop.m 0 C M1 mov b0 = save_rp C I0 ;; }{.mii; nop.m 0 C M0 popcnt x_orig = x_orig C I0 orig x twos popcnt r9 = r9 C I0 x twos ;; }{.mmi; cmp.lt p7,p0 = x_orig, y_twos C M0 orig x_twos < y_twos addl r22 = @@ltoff(ctz_table), r1 shr.u x = x, r9 C I0 x odd ;; }{.mib; (p7) mov y_twos = x_orig C M0 common twos d175 5 a179 4 ;; } mov r25 = m4_lshift(MASK, MAXSHIFT) ld8 r22 = [r22] a180 1 ;; d182 5 a186 4 ALIGN(32) L(top): .pred.rel "mutex", p6,p7 {.mmi; (p7) mov y = x d188 1 a188 3 dep r21 = r19, r22, 0, MAXSHIFT C concat(table,lowbits) }{.mmi; and r20 = MASK, r19 (p7) mov x = r19 d190 2 a191 2 ;; } d193 1 a193 1 {.mmb; ld1 r16 = [r21] d196 3 a198 2 ;; }{.mmi; nop 0 d201 1 a201 2 ;; } d203 1 a203 1 {.mmi; sub r19 = y, x d206 1 a206 1 }{.mmb; nop 0 a208 1 } d210 3 a212 1 L(done_y): C result is y d218 1 a218 1 and r20 = x, r25 a219 2 ;; dep r21 = x, r22, 0, MAXSHIFT @