![[BACK]](/icons/cvsweb/back.gif){kind=icon}
Return to lshift.asm CVS log ![[TXT]](/icons/cvsweb/text.gif){kind=icon}
![[DIR]](/icons/cvsweb/dir.gif){kind=icon}
Up to [local] / OpenXM_contrib / gmp / mpn / x86 / k6 / k62mmx|
Return to lshift.asm CVS log | Up to [local] / OpenXM_contrib / gmp / mpn / x86 / k6 / k62mmx |
1.1 maekawa 1: dnl AMD K6-2 mpn_lshift -- mpn left shift.
2:
1.1.1.2 ! ohara 3: dnl Copyright 1999, 2000, 2002 Free Software Foundation, Inc.
1.1 maekawa 4: dnl
5: dnl This file is part of the GNU MP Library.
6: dnl
7: dnl The GNU MP Library is free software; you can redistribute it and/or
8: dnl modify it under the terms of the GNU Lesser General Public License as
9: dnl published by the Free Software Foundation; either version 2.1 of the
10: dnl License, or (at your option) any later version.
11: dnl
12: dnl The GNU MP Library is distributed in the hope that it will be useful,
13: dnl but WITHOUT ANY WARRANTY; without even the implied warranty of
14: dnl MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15: dnl Lesser General Public License for more details.
16: dnl
17: dnl You should have received a copy of the GNU Lesser General Public
18: dnl License along with the GNU MP Library; see the file COPYING.LIB. If
19: dnl not, write to the Free Software Foundation, Inc., 59 Temple Place -
20: dnl Suite 330, Boston, MA 02111-1307, USA.
21:
22: include(`../config.m4')
23:
24:
1.1.1.2 ! ohara 25: C K6-2: 1.75 cycles/limb
! 26:
! 27:
1.1 maekawa 28: C mp_limb_t mpn_lshift (mp_ptr dst, mp_srcptr src, mp_size_t size,
29: C unsigned shift);
30: C
31:
32: defframe(PARAM_SHIFT,16)
33: defframe(PARAM_SIZE, 12)
34: defframe(PARAM_SRC, 8)
35: defframe(PARAM_DST, 4)
36: deflit(`FRAME',0)
37:
38: dnl used after src has been fetched
39: define(VAR_RETVAL,`PARAM_SRC')
40:
41: dnl minimum 9, because unrolled loop can't handle less
42: deflit(UNROLL_THRESHOLD, 9)
43:
1.1.1.2 ! ohara 44: TEXT
1.1 maekawa 45: ALIGN(32)
46:
47: PROLOGUE(mpn_lshift)
48: deflit(`FRAME',0)
49:
50: C The 1 limb case can be done without the push %ebx, but it's then
51: C still the same speed. The push is left as a free helping hand for
52: C the two_or_more code.
53:
54: movl PARAM_SIZE, %eax
55: pushl %ebx FRAME_pushl()
56:
57: movl PARAM_SRC, %ebx
58: decl %eax
59:
60: movl PARAM_SHIFT, %ecx
61: jnz L(two_or_more)
62:
63: movl (%ebx), %edx C src limb
64: movl PARAM_DST, %ebx
65:
66: shldl( %cl, %edx, %eax) C return value
67:
68: shll %cl, %edx
69:
70: movl %edx, (%ebx) C dst limb
71: popl %ebx
72:
73: ret
74:
75:
76: C -----------------------------------------------------------------------------
77: ALIGN(16) C avoid offset 0x1f
78: L(two_or_more):
79: C eax size-1
80: C ebx src
81: C ecx shift
82: C edx
83:
84: movl (%ebx,%eax,4), %edx C src high limb
85: negl %ecx
86:
87: movd PARAM_SHIFT, %mm6
88: addl $32, %ecx C 32-shift
89:
90: shrl %cl, %edx
91: cmpl $UNROLL_THRESHOLD-1, %eax
92:
93: movl %edx, VAR_RETVAL
94: jae L(unroll)
95:
96:
97: movd %ecx, %mm7
98: movl %eax, %ecx
99:
100: movl PARAM_DST, %eax
101:
102: L(simple):
103: C eax dst
104: C ebx src
105: C ecx counter, size-1 to 1
106: C edx retval
107: C
108: C mm0 scratch
109: C mm6 shift
110: C mm7 32-shift
111:
112: movq -4(%ebx,%ecx,4), %mm0
113:
114: psrlq %mm7, %mm0
115:
116: Zdisp( movd, %mm0, 0,(%eax,%ecx,4))
117: loop L(simple)
118:
119:
120: movd (%ebx), %mm0
121: popl %ebx
122:
123: psllq %mm6, %mm0
124:
125: movd %mm0, (%eax)
126: movl %edx, %eax
127:
128: femms
129: ret
130:
131:
132: C -----------------------------------------------------------------------------
133: ALIGN(16)
134: L(unroll):
135: C eax size-1
136: C ebx src
137: C ecx 32-shift
138: C edx retval (but instead VAR_RETVAL is used)
139: C
140: C mm6 shift
141:
142: addl $32, %ecx
143: movl PARAM_DST, %edx
144:
145: movd %ecx, %mm7
146: subl $7, %eax C size-8
147:
148: leal (%edx,%eax,4), %ecx C alignment of dst
149:
150: movq 32-8(%ebx,%eax,4), %mm2 C src high qword
151: testb $4, %cl
152:
153: jz L(dst_aligned)
154: psllq %mm6, %mm2
155:
156: psrlq $32, %mm2
157: decl %eax
158:
159: movd %mm2, 32(%edx,%eax,4) C dst high limb
160: movq 32-8(%ebx,%eax,4), %mm2 C new src high qword
161: L(dst_aligned):
162:
163: movq 32-16(%ebx,%eax,4), %mm0 C src second highest qword
164:
165:
166: C This loop is the important bit, the rest is just support for it.
167: C Four src limbs are held at the start, and four more will be read.
168: C Four dst limbs will be written. This schedule seems necessary for
169: C full speed.
170: C
171: C The use of size-8 lets the loop stop when %eax goes negative and
172: C leaves -4 to -1 which can be tested with test $1 and $2.
173:
174: L(top):
175: C eax counter, size-8 step by -4 until <0
176: C ebx src
177: C ecx
178: C edx dst
179: C
180: C mm0 src next qword
181: C mm1 scratch
182: C mm2 src prev qword
183: C mm6 shift
184: C mm7 64-shift
185:
186: psllq %mm6, %mm2
187: subl $4, %eax
188:
189: movq %mm0, %mm1
190: psrlq %mm7, %mm0
191:
192: por %mm0, %mm2
193: movq 24(%ebx,%eax,4), %mm0
194:
195: psllq %mm6, %mm1
196: movq %mm2, 40(%edx,%eax,4)
197:
198: movq %mm0, %mm2
199: psrlq %mm7, %mm0
200:
201: por %mm0, %mm1
202: movq 16(%ebx,%eax,4), %mm0
203:
204: movq %mm1, 32(%edx,%eax,4)
205: jnc L(top)
206:
207:
208: C Now have four limbs in mm2 (prev) and mm0 (next), plus eax mod 4.
209: C
210: C 8(%ebx) is the next source, and 24(%edx) is the next destination.
211: C %eax is between -4 and -1, representing respectively 0 to 3 extra
212: C limbs that must be read.
213:
214:
215: testl $2, %eax C testl to avoid bad cache line crossing
216: jz L(finish_nottwo)
217:
218: C Two more limbs: lshift mm2, OR it with rshifted mm0, mm0 becomes
219: C new mm2 and a new mm0 is loaded.
220:
221: psllq %mm6, %mm2
222: movq %mm0, %mm1
223:
224: psrlq %mm7, %mm0
225: subl $2, %eax
226:
227: por %mm0, %mm2
228: movq 16(%ebx,%eax,4), %mm0
229:
230: movq %mm2, 32(%edx,%eax,4)
231: movq %mm1, %mm2
232: L(finish_nottwo):
233:
234:
235: C lshift mm2, OR with rshifted mm0, mm1 becomes lshifted mm0
236:
237: testb $1, %al
238: psllq %mm6, %mm2
239:
240: movq %mm0, %mm1
241: psrlq %mm7, %mm0
242:
243: por %mm0, %mm2
244: psllq %mm6, %mm1
245:
246: movq %mm2, 24(%edx,%eax,4)
247: jz L(finish_even)
248:
249:
250: C Size is odd, so mm1 and one extra limb to process.
251:
252: movd (%ebx), %mm0 C src[0]
253: popl %ebx
254: deflit(`FRAME',0)
255:
256: movq %mm0, %mm2
257: psllq $32, %mm0
258:
259: psrlq %mm7, %mm0
260:
261: psllq %mm6, %mm2
262: por %mm0, %mm1
263:
264: movq %mm1, 4(%edx) C dst[1,2]
265: movd %mm2, (%edx) C dst[0]
266:
267: movl VAR_RETVAL, %eax
268:
269: femms
270: ret
271:
272:
273: nop C avoid bad cache line crossing
274: L(finish_even):
275: deflit(`FRAME',4)
276: C Size is even, so only mm1 left to process.
277:
278: movq %mm1, (%edx) C dst[0,1]
279: movl VAR_RETVAL, %eax
280:
281: popl %ebx
282: femms
283: ret
284:
285: EPILOGUE()