Annotation of OpenXM_contrib/gmp/mpn/x86/pentium/mmx/mul_1.asm, Revision 1.1.1.1
1.1 ohara 1: dnl Intel Pentium MMX mpn_mul_1 -- mpn by limb multiplication.
2:
3: dnl Copyright 2000, 2001, 2002 Free Software Foundation, Inc.
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:
25: C cycles/limb
26: C P5: 12.0 for 32-bit multiplier
27: C 7.0 for 16-bit multiplier
28:
29:
30: C mp_limb_t mpn_mul_1 (mp_ptr dst, mp_srcptr src, mp_size_t size,
31: C mp_limb_t multiplier);
32: C
33: C When the multiplier is 16 bits some special case MMX code is used. Small
34: C multipliers might arise reasonably often from mpz_mul_ui etc. If the size
35: C is odd there's roughly a 5 cycle penalty, so times for say size==7 and
36: C size==8 end up being quite close. If src isn't aligned to an 8 byte
37: C boundary then one limb is processed separately with roughly a 5 cycle
38: C penalty, so in that case it's say size==8 and size==9 which are close.
39: C
40: C Alternatives:
41: C
42: C MMX is not believed to be of any use for 32-bit multipliers, since for
43: C instance the current method would just have to be more or less duplicated
44: C for the high and low halves of the multiplier, and would probably
45: C therefore run at about 14 cycles, which is slower than the plain integer
46: C at 12.
47: C
48: C Adding the high and low MMX products using integer code seems best. An
49: C attempt at using paddd and carry bit propagation with pcmpgtd didn't give
50: C any joy. Perhaps something could be done keeping the values signed and
51: C thereby avoiding adjustments to make pcmpgtd into an unsigned compare, or
52: C perhaps not.
53: C
54: C Future:
55: C
56: C An mpn_mul_1c entrypoint would need a double carry out of the low result
57: C limb in the 16-bit code, unless it could be assumed the carry fits in 16
58: C bits, possibly as carry<multiplier, this being true of a big calculation
59: C done piece by piece. But let's worry about that if/when mul_1c is
60: C actually used.
61:
62: defframe(PARAM_MULTIPLIER,16)
63: defframe(PARAM_SIZE, 12)
64: defframe(PARAM_SRC, 8)
65: defframe(PARAM_DST, 4)
66:
67: TEXT
68:
69: ALIGN(8)
70: PROLOGUE(mpn_mul_1)
71: deflit(`FRAME',0)
72:
73: movl PARAM_SIZE, %ecx
74: movl PARAM_SRC, %edx
75:
76: cmpl $1, %ecx
77: jne L(two_or_more)
78:
79: C one limb only
80:
81: movl PARAM_MULTIPLIER, %eax
82: movl PARAM_DST, %ecx
83:
84: mull (%edx)
85:
86: movl %eax, (%ecx)
87: movl %edx, %eax
88:
89: ret
90:
91:
92: L(two_or_more):
93: C eax size
94: C ebx
95: C ecx carry
96: C edx
97: C esi src
98: C edi
99: C ebp
100:
101: pushl %esi FRAME_pushl()
102: pushl %edi FRAME_pushl()
103:
104: movl %edx, %esi C src
105: movl PARAM_DST, %edi
106:
107: movl PARAM_MULTIPLIER, %eax
108: pushl %ebx FRAME_pushl()
109:
110: leal (%esi,%ecx,4), %esi C src end
111: leal (%edi,%ecx,4), %edi C dst end
112:
113: negl %ecx C -size
114:
115: pushl %ebp FRAME_pushl()
116: cmpl $65536, %eax
117:
118: jb L(small)
119:
120:
121: L(big):
122: xorl %ebx, %ebx C carry limb
123: sarl %ecx C -size/2
124:
125: jnc L(top) C with carry flag clear
126:
127:
128: C size was odd, process one limb separately
129:
130: mull 4(%esi,%ecx,8) C m * src[0]
131:
132: movl %eax, 4(%edi,%ecx,8)
133: incl %ecx
134:
135: orl %edx, %ebx C carry limb, and clear carry flag
136:
137:
138: L(top):
139: C eax
140: C ebx carry
141: C ecx counter, negative
142: C edx
143: C esi src end
144: C edi dst end
145: C ebp (scratch carry)
146:
147: adcl $0, %ebx
148: movl (%esi,%ecx,8), %eax
149:
150: mull PARAM_MULTIPLIER
151:
152: movl %edx, %ebp
153: addl %eax, %ebx
154:
155: adcl $0, %ebp
156: movl 4(%esi,%ecx,8), %eax
157:
158: mull PARAM_MULTIPLIER
159:
160: movl %ebx, (%edi,%ecx,8)
161: addl %ebp, %eax
162:
163: movl %eax, 4(%edi,%ecx,8)
164: incl %ecx
165:
166: movl %edx, %ebx
167: jnz L(top)
168:
169:
170: adcl $0, %ebx
171: popl %ebp
172:
173: movl %ebx, %eax
174: popl %ebx
175:
176: popl %edi
177: popl %esi
178:
179: ret
180:
181:
182: L(small):
183: C Special case for 16-bit multiplier.
184: C
185: C eax multiplier
186: C ebx
187: C ecx -size
188: C edx src
189: C esi src end
190: C edi dst end
191: C ebp multiplier
192:
193: C size<3 not supported here. At size==3 we're already a couple of
194: C cycles faster, so there's no threshold as such, just use the MMX
195: C as soon as possible.
196:
197: cmpl $-3, %ecx
198: ja L(big)
199:
200: movd %eax, %mm7 C m
201: pxor %mm6, %mm6 C initial carry word
202:
203: punpcklwd %mm7, %mm7 C m replicated 2 times
204: addl $2, %ecx C -size+2
205:
206: punpckldq %mm7, %mm7 C m replicated 4 times
207: andl $4, %edx C test alignment, clear carry flag
208:
209: movq %mm7, %mm0 C m
210: jz L(small_entry)
211:
212:
213: C Source is unaligned, process one limb separately.
214: C
215: C Plain integer code is used here, since it's smaller and is about
216: C the same 13 cycles as an mmx block would be.
217: C
218: C An "addl $1,%ecx" doesn't clear the carry flag when size==3, hence
219: C the use of separate incl and orl.
220:
221: mull -8(%esi,%ecx,4) C m * src[0]
222:
223: movl %eax, -8(%edi,%ecx,4) C dst[0]
224: incl %ecx C one limb processed
225:
226: movd %edx, %mm6 C initial carry
227:
228: orl %eax, %eax C clear carry flag
229: jmp L(small_entry)
230:
231:
232: C The scheduling here is quite tricky, since so many instructions have
233: C pairing restrictions. In particular the js won't pair with a movd, and
234: C can't be paired with an adc since it wants flags from the inc, so
235: C instructions are rotated to the top of the loop to find somewhere useful
236: C for it.
237: C
238: C Trouble has been taken to avoid overlapping successive loop iterations,
239: C since that would greatly increase the size of the startup and finishup
240: C code. Actually there's probably not much advantage to be had from
241: C overlapping anyway, since the difficulties are mostly with pairing, not
242: C with latencies as such.
243: C
244: C In the comments x represents the src data and m the multiplier (16
245: C bits, but replicated 4 times).
246: C
247: C The m signs calculated in %mm3 are a loop invariant and could be held in
248: C say %mm5, but that would save only one instruction and hence be no faster.
249:
250: L(small_top):
251: C eax l.low, then l.high
252: C ebx (h.low)
253: C ecx counter, -size+2 to 0 or 1
254: C edx (h.high)
255: C esi &src[size]
256: C edi &dst[size]
257: C ebp
258: C
259: C %mm0 (high products)
260: C %mm1 (low products)
261: C %mm2 (adjust for m using x signs)
262: C %mm3 (adjust for x using m signs)
263: C %mm4
264: C %mm5
265: C %mm6 h.low, then carry
266: C %mm7 m replicated 4 times
267:
268: movd %mm6, %ebx C h.low
269: psrlq $32, %mm1 C l.high
270:
271: movd %mm0, %edx C h.high
272: movq %mm0, %mm6 C new c
273:
274: adcl %eax, %ebx
275: incl %ecx
276:
277: movd %mm1, %eax C l.high
278: movq %mm7, %mm0
279:
280: adcl %eax, %edx
281: movl %ebx, -16(%edi,%ecx,4)
282:
283: movl %edx, -12(%edi,%ecx,4)
284: psrlq $32, %mm6 C c
285:
286: L(small_entry):
287: pmulhw -8(%esi,%ecx,4), %mm0 C h = (x*m).high
288: movq %mm7, %mm1
289:
290: pmullw -8(%esi,%ecx,4), %mm1 C l = (x*m).low
291: movq %mm7, %mm3
292:
293: movq -8(%esi,%ecx,4), %mm2 C x
294: psraw $15, %mm3 C m signs
295:
296: pand -8(%esi,%ecx,4), %mm3 C x selected by m signs
297: psraw $15, %mm2 C x signs
298:
299: paddw %mm3, %mm0 C add x to h if m neg
300: pand %mm7, %mm2 C m selected by x signs
301:
302: paddw %mm2, %mm0 C add m to h if x neg
303: incl %ecx
304:
305: movd %mm1, %eax C l.low
306: punpcklwd %mm0, %mm6 C c + h.low << 16
307:
308: psrlq $16, %mm0 C h.high
309: js L(small_top)
310:
311:
312:
313:
314: movd %mm6, %ebx C h.low
315: psrlq $32, %mm1 C l.high
316:
317: adcl %eax, %ebx
318: popl %ebp FRAME_popl()
319:
320: movd %mm0, %edx C h.high
321: psrlq $32, %mm0 C l.high
322:
323: movd %mm1, %eax C l.high
324:
325: adcl %eax, %edx
326: movl %ebx, -12(%edi,%ecx,4)
327:
328: movd %mm0, %eax C c
329:
330: adcl $0, %eax
331: movl %edx, -8(%edi,%ecx,4)
332:
333: orl %ecx, %ecx
334: jnz L(small_done) C final %ecx==1 means even, ==0 odd
335:
336:
337: C Size odd, one extra limb to process.
338: C Plain integer code is used here, since it's smaller and is about
339: C the same speed as another mmx block would be.
340:
341: movl %eax, %ecx
342: movl PARAM_MULTIPLIER, %eax
343:
344: mull -4(%esi)
345:
346: addl %ecx, %eax
347:
348: adcl $0, %edx
349: movl %eax, -4(%edi)
350:
351: movl %edx, %eax
352: L(small_done):
353: popl %ebx
354:
355: popl %edi
356: popl %esi
357:
358: emms
359:
360: ret
361:
362: EPILOGUE()
FreeBSD-CVSweb <freebsd-cvsweb@FreeBSD.org>