[Gmp-commit] /var/hg/gmp: 3 new changesets
mercurial at gmplib.org
mercurial at gmplib.org
Fri Sep 20 18:15:02 CEST 2013
details: /var/hg/gmp/rev/712db07a7e22
changeset: 16004:712db07a7e22
user: Torbjorn Granlund <tege at gmplib.org>
date: Fri Sep 20 14:16:16 2013 +0200
description:
Postpone pushes, short-circuit a branch.
details: /var/hg/gmp/rev/d3d9c9ce63bc
changeset: 16005:d3d9c9ce63bc
user: Torbjorn Granlund <tege at gmplib.org>
date: Fri Sep 20 14:17:11 2013 +0200
description:
Provide mullo_basecase for Conroe, Wolfdale, Nehalem, Westmere.
details: /var/hg/gmp/rev/927fe93f3210
changeset: 16006:927fe93f3210
user: Torbjorn Granlund <tege at gmplib.org>
date: Fri Sep 20 14:31:22 2013 +0200
description:
Complete rewrite.
diffstat:
mpn/x86_64/core2/mullo_basecase.asm | 416 ++++++++++++++++++
mpn/x86_64/coreisbr/mullo_basecase.asm | 37 +-
mpn/x86_64/k8/redc_1.asm | 754 +++++++++++++++++++++-----------
3 files changed, 943 insertions(+), 264 deletions(-)
diffs (truncated from 1334 to 300 lines):
diff -r 18e377fa1fc7 -r 927fe93f3210 mpn/x86_64/core2/mullo_basecase.asm
--- /dev/null Thu Jan 01 00:00:00 1970 +0000
+++ b/mpn/x86_64/core2/mullo_basecase.asm Fri Sep 20 14:31:22 2013 +0200
@@ -0,0 +1,416 @@
+dnl AMD64 mpn_mullo_basecase optimised for Conroe/Wolfdale/Nehalem/Westmere.
+
+dnl Contributed to the GNU project by Torbjörn Granlund.
+
+dnl Copyright 2008, 2009, 2011, 2012, 2013 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/limb mul_2 addmul_2
+C AMD K8,K9
+C AMD K10
+C AMD bull
+C AMD pile
+C AMD steam
+C AMD bobcat
+C AMD jaguar
+C Intel P4
+C Intel core 4.0 4.18-4.25
+C Intel NHM 3.75 4.06-4.2
+C Intel SBR
+C Intel IBR
+C Intel HWL
+C Intel BWL
+C Intel atom
+C VIA nano
+
+C The inner loops of this code are the result of running a code generation and
+C optimisation tool suite written by David Harvey and Torbjörn Granlund.
+
+C TODO
+C * Implement proper cor2, replacing current cor0.
+C * Offset n by 2 in order to avoid the outer loop cmp. (And sqr_basecase?)
+C * Micro-optimise.
+
+C When playing with pointers, set this to $2 to fall back to conservative
+C indexing in wind-down code.
+define(`I',`$1')
+
+define(`rp', `%rdi')
+define(`up', `%rsi')
+define(`vp_param', `%rdx')
+define(`n_param', `%rcx')
+
+define(`v0', `%r10')
+define(`v1', `%r11')
+define(`w0', `%rbx')
+define(`w1', `%rcx')
+define(`w2', `%rbp')
+define(`w3', `%r12')
+define(`n', `%r9')
+define(`i', `%r13')
+define(`vp', `%r8')
+
+define(`X0', `%r14')
+define(`X1', `%r15')
+
+C rax rbx rcx rdx rdi rsi rbp r8 r9 r10 r11 r12 r13 r14 r15
+
+ABI_SUPPORT(DOS64)
+ABI_SUPPORT(STD64)
+
+define(`ALIGNx', `ALIGN(16)')
+
+define(`N', 85)
+ifdef(`N',,`define(`N',0)')
+define(`MOV', `ifelse(eval(N & $3),0,`mov $1, $2',`lea ($1), $2')')
+
+ASM_START()
+ TEXT
+ ALIGN(32)
+PROLOGUE(mpn_mullo_basecase)
+ FUNC_ENTRY(4)
+
+ mov (up), %rax
+ mov vp_param, vp
+
+ cmp $4, n_param
+ jb L(small)
+
+ mov (vp_param), v0
+ push %rbx
+ lea (rp,n_param,8), rp C point rp at R[un]
+ push %rbp
+ lea (up,n_param,8), up C point up right after U's end
+ push %r12
+ mov $0, R32(n) C FIXME
+ sub n_param, n
+ push %r13
+ mul v0
+ mov 8(vp), v1
+
+ test $1, R8(n_param)
+ jnz L(m2x1)
+
+L(m2x0):test $2, R8(n_param)
+ jnz L(m2b2)
+
+L(m2b0):lea (n), i
+ mov %rax, (rp,n,8)
+ mov %rdx, w1
+ mov (up,n,8), %rax
+ xor R32(w2), R32(w2)
+ jmp L(m2e0)
+
+L(m2b2):lea -2(n), i
+ mov %rax, w2
+ mov (up,n,8), %rax
+ mov %rdx, w3
+ xor R32(w0), R32(w0)
+ jmp L(m2e2)
+
+L(m2x1):test $2, R8(n_param)
+ jnz L(m2b3)
+
+L(m2b1):lea 1(n), i
+ mov %rax, (rp,n,8)
+ mov (up,n,8), %rax
+ mov %rdx, w0
+ xor R32(w1), R32(w1)
+ jmp L(m2e1)
+
+L(m2b3):lea -1(n), i
+ xor R32(w3), R32(w3)
+ mov %rax, w1
+ mov %rdx, w2
+ mov (up,n,8), %rax
+ jmp L(m2e3)
+
+ ALIGNx
+L(m2tp):mul v0
+ add %rax, w3
+ mov -8(up,i,8), %rax
+ mov w3, -8(rp,i,8)
+ adc %rdx, w0
+ adc $0, R32(w1)
+L(m2e1):mul v1
+ add %rax, w0
+ adc %rdx, w1
+ mov $0, R32(w2)
+ mov (up,i,8), %rax
+ mul v0
+ add %rax, w0
+ mov w0, (rp,i,8)
+ adc %rdx, w1
+ mov (up,i,8), %rax
+ adc $0, R32(w2)
+L(m2e0):mul v1
+ add %rax, w1
+ adc %rdx, w2
+ mov 8(up,i,8), %rax
+ mul v0
+ mov $0, R32(w3)
+ add %rax, w1
+ adc %rdx, w2
+ adc $0, R32(w3)
+ mov 8(up,i,8), %rax
+L(m2e3):mul v1
+ add %rax, w2
+ mov w1, 8(rp,i,8)
+ adc %rdx, w3
+ mov $0, R32(w0)
+ mov 16(up,i,8), %rax
+ mul v0
+ add %rax, w2
+ mov 16(up,i,8), %rax
+ adc %rdx, w3
+ adc $0, R32(w0)
+L(m2e2):mul v1
+ mov $0, R32(w1) C FIXME: dead in last iteration
+ add %rax, w3
+ mov 24(up,i,8), %rax
+ mov w2, 16(rp,i,8)
+ adc %rdx, w0 C FIXME: dead in last iteration
+ add $4, i
+ js L(m2tp)
+
+L(m2ed):imul v0, %rax
+ add w3, %rax
+ mov %rax, I(-8(rp),-8(rp,i,8))
+
+ add $2, n
+ lea 16(vp), vp
+ lea -16(up), up
+ cmp $-2, n
+ jge L(cor1)
+
+ push %r14
+ push %r15
+
+L(outer):
+ mov (vp), v0
+ mov 8(vp), v1
+ mov (up,n,8), %rax
+ mul v0
+ test $1, R8(n)
+ jnz L(a1x1)
+
+L(a1x0):mov %rax, X1
+ MOV( %rdx, X0, 8)
+ mov (up,n,8), %rax
+ mul v1
+ test $2, R8(n)
+ jnz L(a110)
+
+L(a100):lea (n), i
+ mov (rp,n,8), w3
+ mov %rax, w0
+ MOV( %rdx, w1, 16)
+ jmp L(lo0)
+
+L(a110):lea 2(n), i
+ mov (rp,n,8), w1
+ mov %rax, w2
+ mov 8(up,n,8), %rax
+ MOV( %rdx, w3, 1)
+ jmp L(lo2)
+
+L(a1x1):mov %rax, X0
+ MOV( %rdx, X1, 2)
+ mov (up,n,8), %rax
+ mul v1
+ test $2, R8(n)
+ jz L(a111)
+
+L(a101):lea 1(n), i
+ MOV( %rdx, w0, 4)
+ mov (rp,n,8), w2
+ mov %rax, w3
+ jmp L(lo1)
+
+L(a111):lea -1(n), i
+ MOV( %rdx, w2, 64)
+ mov %rax, w1
+ mov (rp,n,8), w0
+ mov 8(up,n,8), %rax
+ jmp L(lo3)
+
+ ALIGNx
+L(top): mul v1
+ add w0, w1
+ adc %rax, w2
+ mov -8(up,i,8), %rax
+ MOV( %rdx, w3, 1)
+ adc $0, w3
+L(lo2): mul v0
+ add w1, X1
+ mov X1, -16(rp,i,8)
+ adc %rax, X0
+ MOV( %rdx, X1, 2)
+ adc $0, X1
+ mov -8(up,i,8), %rax
+ mul v1
+ MOV( %rdx, w0, 4)
+ mov -8(rp,i,8), w1
+ add w1, w2
+ adc %rax, w3
+ adc $0, w0
+L(lo1): mov (up,i,8), %rax
+ mul v0
+ add w2, X0
+ adc %rax, X1
+ mov X0, -8(rp,i,8)
+ MOV( %rdx, X0, 8)
+ adc $0, X0
+ mov (up,i,8), %rax
+ mov (rp,i,8), w2
+ mul v1
+ add w2, w3
+ adc %rax, w0
+ MOV( %rdx, w1, 16)
+ adc $0, w1
+L(lo0): mov 8(up,i,8), %rax
+ mul v0
+ add w3, X1
+ mov X1, (rp,i,8)
+ adc %rax, X0
+ MOV( %rdx, X1, 32)
+ mov 8(rp,i,8), w3
+ adc $0, X1
+ mov 8(up,i,8), %rax
+ mul v1
More information about the gmp-commit
mailing list