dnl X86-32 and X86-64 mpn_popcount using SSE2. dnl Copyright 2006, 2007, 2011 Free Software Foundation, Inc. dnl 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 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 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 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 32-bit popcount hamdist C cycles/limb cycles/limb C P5 - C P6 model 0-8,10-12 - C P6 model 9 (Banias) ? C P6 model 13 (Dothan) 4 C P4 model 0 (Willamette) ? C P4 model 1 (?) ? C P4 model 2 (Northwood) 3.9 C P4 model 3 (Prescott) ? C P4 model 4 (Nocona) ? C AMD K6 - C AMD K7 - C AMD K8 ? C 64-bit popcount hamdist C cycles/limb cycles/limb C P4 model 4 (Nocona): 8 C AMD K8,K9 7.5 C AMD K10 3.5 C Intel core2 3.68 C Intel corei 3.15 C Intel atom 10.8 C VIA nano 6.5 C TODO C * Make a mpn_hamdist based on this. Alignment could either be handled by C using movdqu for one operand and movdqa for the other, or by painfully C shifting as we go. Unfortunately, there seem to be no useable shift C instruction, except for one that takes an immediate count. C * It would probably be possible to cut a few cycles/limb using software C pipelining. C * There are 35 decode slots unused by the SSE2 instructions. Loop control C needs just 2 or 3 slots, leaving around 32 slots. This allows a parallel C integer based popcount. Such a combined loop would handle 6 limbs in C about 30 cycles on K8. C * We could save a byte or two by using 32-bit operations on areg. C * Check if using movdqa to a temp of and then register-based pand is faster. ifelse(GMP_LIMB_BITS,`32', ` define(`up', `%edx') define(`n', `%ecx') define(`areg',`%eax') define(`breg',`%ebx') define(`zero',`%xmm4') define(`LIMB32',` $1') define(`LIMB64',`dnl') ',` define(`up', `%rdi') define(`n', `%rsi') define(`areg',`%rax') define(`breg',`%rdx') define(`zero',`%xmm8') define(`LIMB32',`dnl') define(`LIMB64',` $1') ') define(`mm01010101',`%xmm6') define(`mm00110011',`%xmm7') define(`mm00001111',`%xmm2') define(`GMP_LIMB_BYTES', eval(GMP_LIMB_BITS/8)) define(`LIMBS_PER_XMM', eval(16/GMP_LIMB_BYTES)) define(`LIMBS_PER_2XMM', eval(32/GMP_LIMB_BYTES)) undefine(`psadbw') C override inherited m4 version ASM_START() C Make cnsts global to work around Apple relocation bug. ifdef(`DARWIN',` define(`cnsts', MPN(popccnsts)) GLOBL cnsts') TEXT ALIGN(32) PROLOGUE(mpn_popcount) LIMB32(`mov 4(%esp), up ') LIMB32(`mov 8(%esp), n ') LIMB32(`push %ebx ') pxor %xmm3, %xmm3 C zero grand total count LIMB64(`pxor zero, zero ') ifdef(`PIC',` LEA( cnsts, breg) ',` LIMB32(`mov $cnsts, breg ') LIMB64(`movabs $cnsts, breg ') ') movdqa -48(breg), mm01010101 movdqa -32(breg), mm00110011 movdqa -16(breg), mm00001111 mov up, areg and $-16, up C round `up' down to 128-bit boundary and $12, areg C 32:areg = 0, 4, 8, 12 C 64:areg = 0, 8 movdqa (up), %xmm0 pand 64(breg,areg,4), %xmm0 shr $m4_log2(GMP_LIMB_BYTES), %eax add areg, n C compensate n for rounded down `up' pxor %xmm4, %xmm4 sub $LIMBS_PER_XMM, n jbe L(sum) sub $LIMBS_PER_XMM, n ja L(ent) jmp L(lsum) ALIGN(16) L(top): movdqa (up), %xmm0 L(ent): movdqa 16(up), %xmm4 movdqa %xmm0, %xmm1 movdqa %xmm4, %xmm5 psrld $1, %xmm0 psrld $1, %xmm4 pand mm01010101, %xmm0 pand mm01010101, %xmm4 psubd %xmm0, %xmm1 psubd %xmm4, %xmm5 movdqa %xmm1, %xmm0 movdqa %xmm5, %xmm4 psrlq $2, %xmm1 psrlq $2, %xmm5 pand mm00110011, %xmm0 pand mm00110011, %xmm4 pand mm00110011, %xmm1 pand mm00110011, %xmm5 paddq %xmm0, %xmm1 paddq %xmm4, %xmm5 LIMB32(`pxor zero, zero ') add $32, up sub $LIMBS_PER_2XMM, n paddq %xmm5, %xmm1 movdqa %xmm1, %xmm0 psrlq $4, %xmm1 pand mm00001111, %xmm0 pand mm00001111, %xmm1 paddq %xmm0, %xmm1 psadbw zero, %xmm1 paddq %xmm1, %xmm3 C add to grand total jnc L(top) L(end): add $LIMBS_PER_2XMM, n jz L(rt) movdqa (up), %xmm0 pxor %xmm4, %xmm4 sub $LIMBS_PER_XMM, n jbe L(sum) L(lsum): movdqa %xmm0, %xmm4 movdqa 16(up), %xmm0 L(sum): shl $m4_log2(GMP_LIMB_BYTES), n and $12, n pand (breg,n,4), %xmm0 movdqa %xmm0, %xmm1 movdqa %xmm4, %xmm5 psrld $1, %xmm0 psrld $1, %xmm4 pand mm01010101, %xmm0 pand mm01010101, %xmm4 psubd %xmm0, %xmm1 psubd %xmm4, %xmm5 movdqa %xmm1, %xmm0 movdqa %xmm5, %xmm4 psrlq $2, %xmm1 psrlq $2, %xmm5 pand mm00110011, %xmm0 pand mm00110011, %xmm4 pand mm00110011, %xmm1 pand mm00110011, %xmm5 paddq %xmm0, %xmm1 paddq %xmm4, %xmm5 LIMB32(`pxor zero, zero ') paddq %xmm5, %xmm1 movdqa %xmm1, %xmm0 psrlq $4, %xmm1 pand mm00001111, %xmm0 pand mm00001111, %xmm1 paddq %xmm0, %xmm1 psadbw zero, %xmm1 paddq %xmm1, %xmm3 C add to grand total C Add the two 64-bit halves of the grand total counter L(rt): movdqa %xmm3, %xmm0 psrldq $8, %xmm3 paddq %xmm3, %xmm0 movd %xmm0, areg C movq avoided due to gas bug LIMB32(`pop %ebx ') ret EPILOGUE() DEF_OBJECT(dummy,16) C Three magic constants used for masking out bits .byte 0x55,0x55,0x55,0x55,0x55,0x55,0x55,0x55 .byte 0x55,0x55,0x55,0x55,0x55,0x55,0x55,0x55 .byte 0x33,0x33,0x33,0x33,0x33,0x33,0x33,0x33 .byte 0x33,0x33,0x33,0x33,0x33,0x33,0x33,0x33 .byte 0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f .byte 0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f cnsts: C Masks for high end of number .byte 0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff .byte 0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff .byte 0xff,0xff,0xff,0xff,0x00,0x00,0x00,0x00 .byte 0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00 .byte 0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff .byte 0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00 .byte 0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff .byte 0xff,0xff,0xff,0xff,0x00,0x00,0x00,0x00 C Masks for low end of number .byte 0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff .byte 0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff .byte 0x00,0x00,0x00,0x00,0xff,0xff,0xff,0xff .byte 0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff .byte 0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00 .byte 0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff .byte 0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00 .byte 0x00,0x00,0x00,0x00,0xff,0xff,0xff,0xff END_OBJECT(dummy)