VirtualBox

source: vbox/trunk/include/iprt/asm-math.h@ 52454

Last change on this file since 52454 was 52454, checked in by vboxsync, 10 years ago

ASMMultU32ByU32DivByU32: Cleaned up mess. Tested on x86 and amd64-msc.
  • Property svn:eol-style set to native
  • Property svn:keywords set to Author Date Id Revision
File size: 12.8 KB
Line 
1/** @file
2 * IPRT - Assembly Routines for Optimizing some Integers Math Operations.
3 */
4
5/*
6 * Copyright (C) 2006-2010 Oracle Corporation
7 *
8 * This file is part of VirtualBox Open Source Edition (OSE), as
9 * available from http://www.virtualbox.org. This file is free software;
10 * you can redistribute it and/or modify it under the terms of the GNU
11 * General Public License (GPL) as published by the Free Software
12 * Foundation, in version 2 as it comes in the "COPYING" file of the
13 * VirtualBox OSE distribution. VirtualBox OSE is distributed in the
14 * hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
15 *
16 * The contents of this file may alternatively be used under the terms
17 * of the Common Development and Distribution License Version 1.0
18 * (CDDL) only, as it comes in the "COPYING.CDDL" file of the
19 * VirtualBox OSE distribution, in which case the provisions of the
20 * CDDL are applicable instead of those of the GPL.
21 *
22 * You may elect to license modified versions of this file under the
23 * terms and conditions of either the GPL or the CDDL or both.
24 */
25
26#ifndef ___iprt_asm_math_h
27#define ___iprt_asm_math_h
28
29#include <iprt/types.h>
30
31#if defined(_MSC_VER) && RT_INLINE_ASM_USES_INTRIN
32# include <intrin.h>
33 /* Emit the intrinsics at all optimization levels. */
34# pragma intrinsic(__emul)
35# pragma intrinsic(__emulu)
36# ifdef RT_ARCH_AMD64
37# pragma intrinsic(_mul128)
38# pragma intrinsic(_umul128)
39# endif
40#endif
41
42
43/** @defgroup grp_rt_asm_math Interger Math Optimizations
44 * @ingroup grp_rt_asm
45 * @{ */
46
47/**
48 * Multiplies two unsigned 32-bit values returning an unsigned 64-bit result.
49 *
50 * @returns u32F1 * u32F2.
51 */
52
53#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_X86)
54DECLASM(uint64_t) ASMMult2xU32RetU64(uint32_t u32F1, uint32_t u32F2);
55#else
56DECLINLINE(uint64_t) ASMMult2xU32RetU64(uint32_t u32F1, uint32_t u32F2)
57{
58# ifdef RT_ARCH_X86
59 uint64_t u64;
60# if RT_INLINE_ASM_GNU_STYLE
61 __asm__ __volatile__("mull %%edx"
62 : "=A" (u64)
63 : "a" (u32F2), "d" (u32F1));
64# elif RT_INLINE_ASM_USES_INTRIN
65 u64 = __emulu(u32F1, u32F2);
66# else
67 __asm
68 {
69 mov edx, [u32F1]
70 mov eax, [u32F2]
71 mul edx
72 mov dword ptr [u64], eax
73 mov dword ptr [u64 + 4], edx
74 }
75# endif
76 return u64;
77# else /* generic: */
78 return (uint64_t)u32F1 * u32F2;
79# endif
80}
81#endif
82
83
84/**
85 * Multiplies two signed 32-bit values returning a signed 64-bit result.
86 *
87 * @returns u32F1 * u32F2.
88 */
89#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_X86)
90DECLASM(int64_t) ASMMult2xS32RetS64(int32_t i32F1, int32_t i32F2);
91#else
92DECLINLINE(int64_t) ASMMult2xS32RetS64(int32_t i32F1, int32_t i32F2)
93{
94# ifdef RT_ARCH_X86
95 int64_t i64;
96# if RT_INLINE_ASM_GNU_STYLE
97 __asm__ __volatile__("imull %%edx"
98 : "=A" (i64)
99 : "a" (i32F2), "d" (i32F1));
100# elif RT_INLINE_ASM_USES_INTRIN
101 i64 = __emul(i32F1, i32F2);
102# else
103 __asm
104 {
105 mov edx, [i32F1]
106 mov eax, [i32F2]
107 imul edx
108 mov dword ptr [i64], eax
109 mov dword ptr [i64 + 4], edx
110 }
111# endif
112 return i64;
113# else /* generic: */
114 return (int64_t)i32F1 * i32F2;
115# endif
116}
117#endif
118
119
120#if ARCH_BITS == 64
121DECLINLINE(uint64_t) ASMMult2xU64Ret2xU64(uint64_t u64F1, uint64_t u64F2, uint64_t *pu64ProdHi)
122{
123# if defined(RT_ARCH_AMD64) && (RT_INLINE_ASM_GNU_STYLE || RT_INLINE_ASM_USES_INTRIN)
124# if RT_INLINE_ASM_GNU_STYLE
125 uint64_t u64Low, u64High;
126 __asm__ __volatile__("mulq %%rdx"
127 : "=a" (u64Low), "=d" (u64High)
128 : "0" (u64F1), "1" (u64F2));
129 *pu64ProdHi = u64High;
130 return u64Low;
131# elif RT_INLINE_ASM_USES_INTRIN
132 return _umul128(u64F1, u64F2, pu64ProdHi);
133# else
134# error "hmm"
135# endif
136# else /* generic: */
137 /*
138 * F1 * F2 = Prod
139 * -- --
140 * ab * cd = b*d + a*d*10 + b*c*10 + a*c*100
141 *
142 * Where a, b, c and d are 'digits', and 10 is max digit + 1.
143 *
144 * Our digits are 32-bit wide, so instead of 10 we multiply by 4G.
145 * Prod = F1.s.Lo*F2.s.Lo + F1.s.Hi*F2.s.Lo*4G
146 * + F1.s.Lo*F2.s.Hi*4G + F1.s.Hi*F2.s.Hi*4G*4G
147 */
148 RTUINT128U Prod;
149 RTUINT64U Tmp1;
150 uint64_t u64Tmp;
151 RTUINT64U F1, F2;
152 F1.u = u64F1;
153 F2.u = u64F2;
154
155 Prod.s.Lo = ASMMult2xU32RetU64(F1.s.Lo, F2.s.Lo);
156
157 Tmp1.u = ASMMult2xU32RetU64(F1.s.Hi, F2.s.Lo);
158 u64Tmp = (uint64_t)Prod.DWords.dw1 + Tmp1.s.Lo;
159 Prod.DWords.dw1 = (uint32_t)u64Tmp;
160 Prod.s.Hi = Tmp1.s.Hi;
161 Prod.s.Hi += u64Tmp >> 32; /* carry */
162
163 Tmp1.u = ASMMult2xU32RetU64(F1.s.Lo, F2.s.Hi);
164 u64Tmp = (uint64_t)Prod.DWords.dw1 + Tmp1.s.Lo;
165 Prod.DWords.dw1 = (uint32_t)u64Tmp;
166 u64Tmp >>= 32; /* carry */
167 u64Tmp += Prod.DWords.dw2;
168 u64Tmp += Tmp1.s.Hi;
169 Prod.DWords.dw2 = (uint32_t)u64Tmp;
170 Prod.DWords.dw3 += u64Tmp >> 32; /* carry */
171
172 Prod.s.Hi += ASMMult2xU32RetU64(F1.s.Hi, F2.s.Hi);
173 *pu64ProdHi = Prod.s.Hi;
174 return Prod.s.Lo;
175# endif
176}
177#endif
178
179
180
181/**
182 * Divides a 64-bit unsigned by a 32-bit unsigned returning an unsigned 32-bit result.
183 *
184 * @returns u64 / u32.
185 */
186#if RT_INLINE_ASM_EXTERNAL && defined(RT_ARCH_X86)
187DECLASM(uint32_t) ASMDivU64ByU32RetU32(uint64_t u64, uint32_t u32);
188#else
189DECLINLINE(uint32_t) ASMDivU64ByU32RetU32(uint64_t u64, uint32_t u32)
190{
191# ifdef RT_ARCH_X86
192# if RT_INLINE_ASM_GNU_STYLE
193 RTCCUINTREG uDummy;
194 __asm__ __volatile__("divl %3"
195 : "=a" (u32), "=d"(uDummy)
196 : "A" (u64), "r" (u32));
197# else
198 __asm
199 {
200 mov eax, dword ptr [u64]
201 mov edx, dword ptr [u64 + 4]
202 mov ecx, [u32]
203 div ecx
204 mov [u32], eax
205 }
206# endif
207 return u32;
208# else /* generic: */
209 return (uint32_t)(u64 / u32);
210# endif
211}
212#endif
213
214
215/**
216 * Divides a 64-bit signed by a 32-bit signed returning a signed 32-bit result.
217 *
218 * @returns u64 / u32.
219 */
220#if RT_INLINE_ASM_EXTERNAL && defined(RT_ARCH_X86)
221DECLASM(int32_t) ASMDivS64ByS32RetS32(int64_t i64, int32_t i32);
222#else
223DECLINLINE(int32_t) ASMDivS64ByS32RetS32(int64_t i64, int32_t i32)
224{
225# ifdef RT_ARCH_X86
226# if RT_INLINE_ASM_GNU_STYLE
227 RTCCUINTREG iDummy;
228 __asm__ __volatile__("idivl %3"
229 : "=a" (i32), "=d"(iDummy)
230 : "A" (i64), "r" (i32));
231# else
232 __asm
233 {
234 mov eax, dword ptr [i64]
235 mov edx, dword ptr [i64 + 4]
236 mov ecx, [i32]
237 idiv ecx
238 mov [i32], eax
239 }
240# endif
241 return i32;
242# else /* generic: */
243 return (int32_t)(i64 / i32);
244# endif
245}
246#endif
247
248
249/**
250 * Performs 64-bit unsigned by a 32-bit unsigned division with a 32-bit unsigned result,
251 * returning the rest.
252 *
253 * @returns u64 % u32.
254 *
255 * @remarks It is important that the result is <= UINT32_MAX or we'll overflow and crash.
256 */
257#if RT_INLINE_ASM_EXTERNAL && defined(RT_ARCH_X86)
258DECLASM(uint32_t) ASMModU64ByU32RetU32(uint64_t u64, uint32_t u32);
259#else
260DECLINLINE(uint32_t) ASMModU64ByU32RetU32(uint64_t u64, uint32_t u32)
261{
262# ifdef RT_ARCH_X86
263# if RT_INLINE_ASM_GNU_STYLE
264 RTCCUINTREG uDummy;
265 __asm__ __volatile__("divl %3"
266 : "=a" (uDummy), "=d"(u32)
267 : "A" (u64), "r" (u32));
268# else
269 __asm
270 {
271 mov eax, dword ptr [u64]
272 mov edx, dword ptr [u64 + 4]
273 mov ecx, [u32]
274 div ecx
275 mov [u32], edx
276 }
277# endif
278 return u32;
279# else /* generic: */
280 return (uint32_t)(u64 % u32);
281# endif
282}
283#endif
284
285
286/**
287 * Performs 64-bit signed by a 32-bit signed division with a 32-bit signed result,
288 * returning the rest.
289 *
290 * @returns u64 % u32.
291 *
292 * @remarks It is important that the result is <= UINT32_MAX or we'll overflow and crash.
293 */
294#if RT_INLINE_ASM_EXTERNAL && defined(RT_ARCH_X86)
295DECLASM(int32_t) ASMModS64ByS32RetS32(int64_t i64, int32_t i32);
296#else
297DECLINLINE(int32_t) ASMModS64ByS32RetS32(int64_t i64, int32_t i32)
298{
299# ifdef RT_ARCH_X86
300# if RT_INLINE_ASM_GNU_STYLE
301 RTCCUINTREG iDummy;
302 __asm__ __volatile__("idivl %3"
303 : "=a" (iDummy), "=d"(i32)
304 : "A" (i64), "r" (i32));
305# else
306 __asm
307 {
308 mov eax, dword ptr [i64]
309 mov edx, dword ptr [i64 + 4]
310 mov ecx, [i32]
311 idiv ecx
312 mov [i32], edx
313 }
314# endif
315 return i32;
316# else /* generic: */
317 return (int32_t)(i64 % i32);
318# endif
319}
320#endif
321
322
323/**
324 * Multiple a 32-bit by a 32-bit integer and divide the result by a 32-bit integer
325 * using a 64 bit intermediate result.
326 *
327 * @returns (u32A * u32B) / u32C.
328 * @param u32A The 32-bit value (A).
329 * @param u32B The 32-bit value to multiple by A.
330 * @param u32C The 32-bit value to divide A*B by.
331 *
332 * @remarks Architecture specific.
333 */
334#if RT_INLINE_ASM_EXTERNAL && (defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86))
335DECLASM(uint32_t) ASMMultU32ByU32DivByU32(uint32_t u32A, uint32_t u32B, uint32_t u32C);
336#else
337DECLINLINE(uint32_t) ASMMultU32ByU32DivByU32(uint32_t u32A, uint32_t u32B, uint32_t u32C)
338{
339# if RT_INLINE_ASM_GNU_STYLE && (defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86))
340 uint32_t u32Result, u32Spill;
341 __asm__ __volatile__("mull %2\n\t"
342 "divl %3\n\t"
343 : "=&a" (u32Result),
344 "=&d" (u32Spill)
345 : "r" (u32B),
346 "r" (u32C),
347 "0" (u32A),
348 "1" (0));
349 return u32Result;
350# else
351 return (uint32_t)(((uint64_t)u32A * u32B) / u32C);
352# endif
353}
354#endif
355
356
357/**
358 * Multiple a 64-bit by a 32-bit integer and divide the result by a 32-bit integer
359 * using a 96 bit intermediate result.
360 * @note Don't use 64-bit C arithmetic here since some gcc compilers generate references to
361 * __udivdi3 and __umoddi3 even if this inline function is not used.
362 *
363 * @returns (u64A * u32B) / u32C.
364 * @param u64A The 64-bit value.
365 * @param u32B The 32-bit value to multiple by A.
366 * @param u32C The 32-bit value to divide A*B by.
367 *
368 * @remarks Architecture specific.
369 */
370#if RT_INLINE_ASM_EXTERNAL || !defined(__GNUC__) || (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
371DECLASM(uint64_t) ASMMultU64ByU32DivByU32(uint64_t u64A, uint32_t u32B, uint32_t u32C);
372#else
373DECLINLINE(uint64_t) ASMMultU64ByU32DivByU32(uint64_t u64A, uint32_t u32B, uint32_t u32C)
374{
375# if RT_INLINE_ASM_GNU_STYLE
376# ifdef RT_ARCH_AMD64
377 uint64_t u64Result, u64Spill;
378 __asm__ __volatile__("mulq %2\n\t"
379 "divq %3\n\t"
380 : "=&a" (u64Result),
381 "=&d" (u64Spill)
382 : "r" ((uint64_t)u32B),
383 "r" ((uint64_t)u32C),
384 "0" (u64A),
385 "1" (0));
386 return u64Result;
387# else
388 uint32_t u32Dummy;
389 uint64_t u64Result;
390 __asm__ __volatile__("mull %%ecx \n\t" /* eax = u64Lo.lo = (u64A.lo * u32B).lo
391 edx = u64Lo.hi = (u64A.lo * u32B).hi */
392 "xchg %%eax,%%esi \n\t" /* esi = u64Lo.lo
393 eax = u64A.hi */
394 "xchg %%edx,%%edi \n\t" /* edi = u64Low.hi
395 edx = u32C */
396 "xchg %%edx,%%ecx \n\t" /* ecx = u32C
397 edx = u32B */
398 "mull %%edx \n\t" /* eax = u64Hi.lo = (u64A.hi * u32B).lo
399 edx = u64Hi.hi = (u64A.hi * u32B).hi */
400 "addl %%edi,%%eax \n\t" /* u64Hi.lo += u64Lo.hi */
401 "adcl $0,%%edx \n\t" /* u64Hi.hi += carry */
402 "divl %%ecx \n\t" /* eax = u64Hi / u32C
403 edx = u64Hi % u32C */
404 "movl %%eax,%%edi \n\t" /* edi = u64Result.hi = u64Hi / u32C */
405 "movl %%esi,%%eax \n\t" /* eax = u64Lo.lo */
406 "divl %%ecx \n\t" /* u64Result.lo */
407 "movl %%edi,%%edx \n\t" /* u64Result.hi */
408 : "=A"(u64Result), "=c"(u32Dummy),
409 "=S"(u32Dummy), "=D"(u32Dummy)
410 : "a"((uint32_t)u64A),
411 "S"((uint32_t)(u64A >> 32)),
412 "c"(u32B),
413 "D"(u32C));
414 return u64Result;
415# endif
416# else
417 RTUINT64U u;
418 uint64_t u64Lo = (uint64_t)(u64A & 0xffffffff) * u32B;
419 uint64_t u64Hi = (uint64_t)(u64A >> 32) * u32B;
420 u64Hi += (u64Lo >> 32);
421 u.s.Hi = (uint32_t)(u64Hi / u32C);
422 u.s.Lo = (uint32_t)((((u64Hi % u32C) << 32) + (u64Lo & 0xffffffff)) / u32C);
423 return u.u;
424# endif
425}
426#endif
427
428/** @} */
429#endif
430
Note: See TracBrowser for help on using the repository browser.

© 2024 Oracle Support Privacy / Do Not Sell My Info Terms of Use Trademark Policy Automated Access Etiquette