--- /dev/null
+dnl AMD K6-2 mpn_copyi -- copy limb vector, incrementing.
+dnl
+dnl K6-2: 0.56 or 1.0 cycles/limb (at 32 limbs/loop), depending on data
+dnl alignment.
+
+
+dnl Copyright (C) 1999, 2000 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
+dnl modify it under the terms of the GNU Lesser General Public License as
+dnl published by the Free Software Foundation; either version 2.1 of the
+dnl License, or (at your option) any later version.
+dnl
+dnl The GNU MP Library is distributed in the hope that it will be useful,
+dnl but WITHOUT ANY WARRANTY; without even the implied warranty of
+dnl MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+dnl Lesser General Public License for more details.
+dnl
+dnl You should have received a copy of the GNU Lesser General Public
+dnl License along with the GNU MP Library; see the file COPYING.LIB. If
+dnl not, write to the Free Software Foundation, Inc., 59 Temple Place -
+dnl Suite 330, Boston, MA 02111-1307, USA.
+
+
+include(`../config.m4')
+
+
+dnl K6-2 aligned:
+dnl UNROLL_COUNT cycles/limb
+dnl 8 0.75
+dnl 16 0.625
+dnl 32 0.5625
+dnl 64 0.53
+dnl Maximum possible with the current code is 64, the minimum is 2.
+
+deflit(UNROLL_COUNT, 32)
+
+
+C void mpn_copyi (mp_ptr dst, mp_srcptr src, mp_size_t size);
+C
+C The MMX loop is faster than a rep movs when src and dst are both 0mod8.
+C With one 0mod8 and one 4mod8 it's 1.056 c/l and the rep movs at 1.0 c/l is
+C used instead.
+C
+C mod8
+C src dst
+C 0 0 both aligned, use mmx
+C 0 4 unaligned, use rep movs
+C 4 0 unaligned, use rep movs
+C 4 4 do one movs, then both aligned, use mmx
+C
+C The MMX code on aligned data is 0.5 c/l, plus loop overhead of 2
+C cycles/loop, which is 0.0625 c/l at 32 limbs/loop.
+C
+C A pattern of two movq loads and two movq stores (or four and four) was
+C tried, but found to be the same speed as just one of each.
+C
+C Note that this code only suits K6-2 and K6-3. Plain K6 does only one mmx
+C instruction per cycle, so "movq"s are no faster than the simple 1 c/l rep
+C movs.
+C
+C Enhancement:
+C
+C Addressing modes like disp(%esi,%ecx,4) aren't currently used. They'd
+C make it possible to avoid incrementing %esi and %edi in the loop and hence
+C get loop overhead down to 1 cycle. Care would be needed to avoid bad
+C cache line crossings since the "movq"s would then be 5 code bytes rather
+C than 4.
+
+
+defframe(PARAM_SIZE,12)
+defframe(PARAM_SRC, 8)
+defframe(PARAM_DST, 4)
+deflit(`FRAME',0)
+
+ .text
+ ALIGN(32)
+
+PROLOGUE(mpn_copyi)
+ movl PARAM_SIZE, %ecx
+ movl %esi, %eax
+
+ movl PARAM_SRC, %esi
+ movl %edi, %edx
+
+ cld
+
+ movl PARAM_DST, %edi
+ cmpl $UNROLL_COUNT, %ecx
+
+ ja L(unroll)
+
+L(simple):
+ rep
+ movsl
+
+ movl %eax, %esi
+ movl %edx, %edi
+
+ ret
+
+
+L(unroll):
+ C if src and dst are different alignments mod8, then use rep movs
+ C if src and dst are both 4mod8 then process one limb to get 0mod8
+
+ pushl %ebx
+ leal (%esi,%edi), %ebx
+
+ testb $4, %bl
+ popl %ebx
+
+ jnz L(simple)
+ testl $4, %esi
+
+ leal -UNROLL_COUNT(%ecx), %ecx
+ jz L(already_aligned)
+
+ decl %ecx
+
+ movsl
+L(already_aligned):
+
+
+ifelse(UNROLL_BYTES,256,`
+ addl $128, %esi
+ addl $128, %edi
+')
+
+ C this is offset 0x34, no alignment needed
+L(top):
+ C eax saved esi
+ C ebx
+ C ecx counter, limbs
+ C edx saved edi
+ C esi src, incrementing
+ C edi dst, incrementing
+ C ebp
+ C
+ C Zdisp gets 0(%esi) left that way to avoid vector decode, and with
+ C 0(%edi) keeps code aligned to 16 byte boundaries.
+
+deflit(CHUNK_COUNT, 2)
+forloop(`i', 0, UNROLL_COUNT/CHUNK_COUNT-1, `
+ deflit(`disp', eval(i*CHUNK_COUNT*4 ifelse(UNROLL_BYTES,256,-128)))
+Zdisp( movq, disp,(%esi), %mm0)
+Zdisp( movq, %mm0, disp,(%edi))
+')
+
+ addl $UNROLL_BYTES, %esi
+ subl $UNROLL_COUNT, %ecx
+
+ leal UNROLL_BYTES(%edi), %edi
+ jns L(top)
+
+
+ C now %ecx is -UNROLL_COUNT to -1 representing repectively 0 to
+ C UNROLL_COUNT-1 limbs remaining
+
+ testb $eval(UNROLL_COUNT/2), %cl
+
+ leal UNROLL_COUNT(%ecx), %ecx
+ jz L(not_half)
+
+ C at an unroll count of 32 this block of code is 16 cycles faster than
+ C the rep movs, less 3 or 4 to test whether to do it
+
+forloop(`i', 0, UNROLL_COUNT/CHUNK_COUNT/2-1, `
+ deflit(`disp', eval(i*CHUNK_COUNT*4 ifelse(UNROLL_BYTES,256,-128)))
+ movq disp(%esi), %mm0
+ movq %mm0, disp(%edi)
+')
+ addl $eval(UNROLL_BYTES/2), %esi
+ addl $eval(UNROLL_BYTES/2), %edi
+
+ subl $eval(UNROLL_COUNT/2), %ecx
+L(not_half):
+
+
+ifelse(UNROLL_BYTES,256,`
+ subl $128, %esi
+ subl $128, %edi
+')
+
+ rep
+ movsl
+
+ movl %eax, %esi
+ movl %edx, %edi
+
+ femms
+ ret
+
+EPILOGUE()
projects / ghc-hetmet.git / blobdiff