VirtualBox

source: vbox/trunk/include/iprt/asm-math.h@ 81025

Last change on this file since 81025 was 76585, checked in by vboxsync, 6 years ago

*: scm --fix-header-guard-endif
  • Property svn:eol-style set to native
  • Property svn:keywords set to Author Date Id Revision
File size: 13.3 KB
Line 
1/** @file
2 * IPRT - Assembly Routines for Optimizing some Integers Math Operations.
3 */
4
5/*
6 * Copyright (C) 2006-2019 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_INCLUDED_asm_math_h
27#define IPRT_INCLUDED_asm_math_h
28#ifndef RT_WITHOUT_PRAGMA_ONCE
29# pragma once
30#endif
31
32#include <iprt/types.h>
33
34#if defined(_MSC_VER) && RT_INLINE_ASM_USES_INTRIN
35# pragma warning(push)
36# pragma warning(disable:4668) /* Several incorrect __cplusplus uses. */
37# pragma warning(disable:4255) /* Incorrect __slwpcb prototype. */
38# include <intrin.h>
39# pragma warning(pop)
40 /* Emit the intrinsics at all optimization levels. */
41# pragma intrinsic(__emul)
42# pragma intrinsic(__emulu)
43# ifdef RT_ARCH_AMD64
44# pragma intrinsic(_mul128)
45# pragma intrinsic(_umul128)
46# endif
47#endif
48
49
50/** @defgroup grp_rt_asm_math Interger Math Optimizations
51 * @ingroup grp_rt_asm
52 * @{ */
53
54/**
55 * Multiplies two unsigned 32-bit values returning an unsigned 64-bit result.
56 *
57 * @returns u32F1 * u32F2.
58 */
59
60#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_X86)
61DECLASM(uint64_t) ASMMult2xU32RetU64(uint32_t u32F1, uint32_t u32F2);
62#else
63DECLINLINE(uint64_t) ASMMult2xU32RetU64(uint32_t u32F1, uint32_t u32F2)
64{
65# ifdef RT_ARCH_X86
66 uint64_t u64;
67# if RT_INLINE_ASM_GNU_STYLE
68 __asm__ __volatile__("mull %%edx"
69 : "=A" (u64)
70 : "a" (u32F2), "d" (u32F1));
71# elif RT_INLINE_ASM_USES_INTRIN
72 u64 = __emulu(u32F1, u32F2);
73# else
74 __asm
75 {
76 mov edx, [u32F1]
77 mov eax, [u32F2]
78 mul edx
79 mov dword ptr [u64], eax
80 mov dword ptr [u64 + 4], edx
81 }
82# endif
83 return u64;
84# else /* generic: */
85 return (uint64_t)u32F1 * u32F2;
86# endif
87}
88#endif
89
90
91/**
92 * Multiplies two signed 32-bit values returning a signed 64-bit result.
93 *
94 * @returns u32F1 * u32F2.
95 */
96#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_X86)
97DECLASM(int64_t) ASMMult2xS32RetS64(int32_t i32F1, int32_t i32F2);
98#else
99DECLINLINE(int64_t) ASMMult2xS32RetS64(int32_t i32F1, int32_t i32F2)
100{
101# ifdef RT_ARCH_X86
102 int64_t i64;
103# if RT_INLINE_ASM_GNU_STYLE
104 __asm__ __volatile__("imull %%edx"
105 : "=A" (i64)
106 : "a" (i32F2), "d" (i32F1));
107# elif RT_INLINE_ASM_USES_INTRIN
108 i64 = __emul(i32F1, i32F2);
109# else
110 __asm
111 {
112 mov edx, [i32F1]
113 mov eax, [i32F2]
114 imul edx
115 mov dword ptr [i64], eax
116 mov dword ptr [i64 + 4], edx
117 }
118# endif
119 return i64;
120# else /* generic: */
121 return (int64_t)i32F1 * i32F2;
122# endif
123}
124#endif
125
126
127#if ARCH_BITS == 64
128DECLINLINE(uint64_t) ASMMult2xU64Ret2xU64(uint64_t u64F1, uint64_t u64F2, uint64_t *pu64ProdHi)
129{
130# if defined(RT_ARCH_AMD64) && (RT_INLINE_ASM_GNU_STYLE || RT_INLINE_ASM_USES_INTRIN)
131# if RT_INLINE_ASM_GNU_STYLE
132 uint64_t u64Low, u64High;
133 __asm__ __volatile__("mulq %%rdx"
134 : "=a" (u64Low), "=d" (u64High)
135 : "0" (u64F1), "1" (u64F2));
136 *pu64ProdHi = u64High;
137 return u64Low;
138# elif RT_INLINE_ASM_USES_INTRIN
139 return _umul128(u64F1, u64F2, pu64ProdHi);
140# else
141# error "hmm"
142# endif
143# else /* generic: */
144 /*
145 * F1 * F2 = Prod
146 * -- --
147 * ab * cd = b*d + a*d*10 + b*c*10 + a*c*100
148 *
149 * Where a, b, c and d are 'digits', and 10 is max digit + 1.
150 *
151 * Our digits are 32-bit wide, so instead of 10 we multiply by 4G.
152 * Prod = F1.s.Lo*F2.s.Lo + F1.s.Hi*F2.s.Lo*4G
153 * + F1.s.Lo*F2.s.Hi*4G + F1.s.Hi*F2.s.Hi*4G*4G
154 */
155 RTUINT128U Prod;
156 RTUINT64U Tmp1;
157 uint64_t u64Tmp;
158 RTUINT64U F1, F2;
159 F1.u = u64F1;
160 F2.u = u64F2;
161
162 Prod.s.Lo = ASMMult2xU32RetU64(F1.s.Lo, F2.s.Lo);
163
164 Tmp1.u = ASMMult2xU32RetU64(F1.s.Hi, F2.s.Lo);
165 u64Tmp = (uint64_t)Prod.DWords.dw1 + Tmp1.s.Lo;
166 Prod.DWords.dw1 = (uint32_t)u64Tmp;
167 Prod.s.Hi = Tmp1.s.Hi;
168 Prod.s.Hi += u64Tmp >> 32; /* carry */
169
170 Tmp1.u = ASMMult2xU32RetU64(F1.s.Lo, F2.s.Hi);
171 u64Tmp = (uint64_t)Prod.DWords.dw1 + Tmp1.s.Lo;
172 Prod.DWords.dw1 = (uint32_t)u64Tmp;
173 u64Tmp >>= 32; /* carry */
174 u64Tmp += Prod.DWords.dw2;
175 u64Tmp += Tmp1.s.Hi;
176 Prod.DWords.dw2 = (uint32_t)u64Tmp;
177 Prod.DWords.dw3 += u64Tmp >> 32; /* carry */
178
179 Prod.s.Hi += ASMMult2xU32RetU64(F1.s.Hi, F2.s.Hi);
180 *pu64ProdHi = Prod.s.Hi;
181 return Prod.s.Lo;
182# endif
183}
184#endif
185
186
187
188/**
189 * Divides a 64-bit unsigned by a 32-bit unsigned returning an unsigned 32-bit result.
190 *
191 * @returns u64 / u32.
192 */
193#if RT_INLINE_ASM_EXTERNAL && defined(RT_ARCH_X86)
194DECLASM(uint32_t) ASMDivU64ByU32RetU32(uint64_t u64, uint32_t u32);
195#else
196DECLINLINE(uint32_t) ASMDivU64ByU32RetU32(uint64_t u64, uint32_t u32)
197{
198# ifdef RT_ARCH_X86
199# if RT_INLINE_ASM_GNU_STYLE
200 RTCCUINTREG uDummy;
201 __asm__ __volatile__("divl %3"
202 : "=a" (u32), "=d"(uDummy)
203 : "A" (u64), "r" (u32));
204# else
205 __asm
206 {
207 mov eax, dword ptr [u64]
208 mov edx, dword ptr [u64 + 4]
209 mov ecx, [u32]
210 div ecx
211 mov [u32], eax
212 }
213# endif
214 return u32;
215# else /* generic: */
216 return (uint32_t)(u64 / u32);
217# endif
218}
219#endif
220
221
222/**
223 * Divides a 64-bit signed by a 32-bit signed returning a signed 32-bit result.
224 *
225 * @returns u64 / u32.
226 */
227#if RT_INLINE_ASM_EXTERNAL && defined(RT_ARCH_X86)
228DECLASM(int32_t) ASMDivS64ByS32RetS32(int64_t i64, int32_t i32);
229#else
230DECLINLINE(int32_t) ASMDivS64ByS32RetS32(int64_t i64, int32_t i32)
231{
232# ifdef RT_ARCH_X86
233# if RT_INLINE_ASM_GNU_STYLE
234 RTCCUINTREG iDummy;
235 __asm__ __volatile__("idivl %3"
236 : "=a" (i32), "=d"(iDummy)
237 : "A" (i64), "r" (i32));
238# else
239 __asm
240 {
241 mov eax, dword ptr [i64]
242 mov edx, dword ptr [i64 + 4]
243 mov ecx, [i32]
244 idiv ecx
245 mov [i32], eax
246 }
247# endif
248 return i32;
249# else /* generic: */
250 return (int32_t)(i64 / i32);
251# endif
252}
253#endif
254
255
256/**
257 * Performs 64-bit unsigned by a 32-bit unsigned division with a 32-bit unsigned result,
258 * returning the rest.
259 *
260 * @returns u64 % u32.
261 *
262 * @remarks It is important that the result is <= UINT32_MAX or we'll overflow and crash.
263 */
264#if RT_INLINE_ASM_EXTERNAL && defined(RT_ARCH_X86)
265DECLASM(uint32_t) ASMModU64ByU32RetU32(uint64_t u64, uint32_t u32);
266#else
267DECLINLINE(uint32_t) ASMModU64ByU32RetU32(uint64_t u64, uint32_t u32)
268{
269# ifdef RT_ARCH_X86
270# if RT_INLINE_ASM_GNU_STYLE
271 RTCCUINTREG uDummy;
272 __asm__ __volatile__("divl %3"
273 : "=a" (uDummy), "=d"(u32)
274 : "A" (u64), "r" (u32));
275# else
276 __asm
277 {
278 mov eax, dword ptr [u64]
279 mov edx, dword ptr [u64 + 4]
280 mov ecx, [u32]
281 div ecx
282 mov [u32], edx
283 }
284# endif
285 return u32;
286# else /* generic: */
287 return (uint32_t)(u64 % u32);
288# endif
289}
290#endif
291
292
293/**
294 * Performs 64-bit signed by a 32-bit signed division with a 32-bit signed result,
295 * returning the rest.
296 *
297 * @returns u64 % u32.
298 *
299 * @remarks It is important that the result is <= UINT32_MAX or we'll overflow and crash.
300 */
301#if RT_INLINE_ASM_EXTERNAL && defined(RT_ARCH_X86)
302DECLASM(int32_t) ASMModS64ByS32RetS32(int64_t i64, int32_t i32);
303#else
304DECLINLINE(int32_t) ASMModS64ByS32RetS32(int64_t i64, int32_t i32)
305{
306# ifdef RT_ARCH_X86
307# if RT_INLINE_ASM_GNU_STYLE
308 RTCCUINTREG iDummy;
309 __asm__ __volatile__("idivl %3"
310 : "=a" (iDummy), "=d"(i32)
311 : "A" (i64), "r" (i32));
312# else
313 __asm
314 {
315 mov eax, dword ptr [i64]
316 mov edx, dword ptr [i64 + 4]
317 mov ecx, [i32]
318 idiv ecx
319 mov [i32], edx
320 }
321# endif
322 return i32;
323# else /* generic: */
324 return (int32_t)(i64 % i32);
325# endif
326}
327#endif
328
329
330/**
331 * Multiple a 32-bit by a 32-bit integer and divide the result by a 32-bit integer
332 * using a 64 bit intermediate result.
333 *
334 * @returns (u32A * u32B) / u32C.
335 * @param u32A The 32-bit value (A).
336 * @param u32B The 32-bit value to multiple by A.
337 * @param u32C The 32-bit value to divide A*B by.
338 *
339 * @remarks Architecture specific.
340 * @remarks Make sure the result won't ever exceed 32-bit, because hardware
341 * exception may be raised if it does.
342 * @remarks On x86 this may be used to avoid dragging in 64-bit builtin
343 * arithmetics functions.
344 */
345#if RT_INLINE_ASM_EXTERNAL && (defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86))
346DECLASM(uint32_t) ASMMultU32ByU32DivByU32(uint32_t u32A, uint32_t u32B, uint32_t u32C);
347#else
348DECLINLINE(uint32_t) ASMMultU32ByU32DivByU32(uint32_t u32A, uint32_t u32B, uint32_t u32C)
349{
350# if RT_INLINE_ASM_GNU_STYLE && (defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86))
351 uint32_t u32Result, u32Spill;
352 __asm__ __volatile__("mull %2\n\t"
353 "divl %3\n\t"
354 : "=&a" (u32Result),
355 "=&d" (u32Spill)
356 : "r" (u32B),
357 "r" (u32C),
358 "0" (u32A));
359 return u32Result;
360# else
361 return (uint32_t)(((uint64_t)u32A * u32B) / u32C);
362# endif
363}
364#endif
365
366
367/**
368 * Multiple a 64-bit by a 32-bit integer and divide the result by a 32-bit integer
369 * using a 96 bit intermediate result.
370 *
371 * @returns (u64A * u32B) / u32C.
372 * @param u64A The 64-bit value.
373 * @param u32B The 32-bit value to multiple by A.
374 * @param u32C The 32-bit value to divide A*B by.
375 *
376 * @remarks Architecture specific.
377 * @remarks Make sure the result won't ever exceed 64-bit, because hardware
378 * exception may be raised if it does.
379 * @remarks On x86 this may be used to avoid dragging in 64-bit builtin
380 * arithmetics function.
381 */
382#if RT_INLINE_ASM_EXTERNAL || !defined(__GNUC__) || (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
383DECLASM(uint64_t) ASMMultU64ByU32DivByU32(uint64_t u64A, uint32_t u32B, uint32_t u32C);
384#else
385DECLINLINE(uint64_t) ASMMultU64ByU32DivByU32(uint64_t u64A, uint32_t u32B, uint32_t u32C)
386{
387# if RT_INLINE_ASM_GNU_STYLE
388# ifdef RT_ARCH_AMD64
389 uint64_t u64Result, u64Spill;
390 __asm__ __volatile__("mulq %2\n\t"
391 "divq %3\n\t"
392 : "=&a" (u64Result),
393 "=&d" (u64Spill)
394 : "r" ((uint64_t)u32B),
395 "r" ((uint64_t)u32C),
396 "0" (u64A));
397 return u64Result;
398# else
399 uint32_t u32Dummy;
400 uint64_t u64Result;
401 __asm__ __volatile__("mull %%ecx \n\t" /* eax = u64Lo.lo = (u64A.lo * u32B).lo
402 edx = u64Lo.hi = (u64A.lo * u32B).hi */
403 "xchg %%eax,%%esi \n\t" /* esi = u64Lo.lo
404 eax = u64A.hi */
405 "xchg %%edx,%%edi \n\t" /* edi = u64Low.hi
406 edx = u32C */
407 "xchg %%edx,%%ecx \n\t" /* ecx = u32C
408 edx = u32B */
409 "mull %%edx \n\t" /* eax = u64Hi.lo = (u64A.hi * u32B).lo
410 edx = u64Hi.hi = (u64A.hi * u32B).hi */
411 "addl %%edi,%%eax \n\t" /* u64Hi.lo += u64Lo.hi */
412 "adcl $0,%%edx \n\t" /* u64Hi.hi += carry */
413 "divl %%ecx \n\t" /* eax = u64Hi / u32C
414 edx = u64Hi % u32C */
415 "movl %%eax,%%edi \n\t" /* edi = u64Result.hi = u64Hi / u32C */
416 "movl %%esi,%%eax \n\t" /* eax = u64Lo.lo */
417 "divl %%ecx \n\t" /* u64Result.lo */
418 "movl %%edi,%%edx \n\t" /* u64Result.hi */
419 : "=A"(u64Result), "=c"(u32Dummy),
420 "=S"(u32Dummy), "=D"(u32Dummy)
421 : "a"((uint32_t)u64A),
422 "S"((uint32_t)(u64A >> 32)),
423 "c"(u32B),
424 "D"(u32C));
425 return u64Result;
426# endif
427# else
428 RTUINT64U u;
429 uint64_t u64Lo = (uint64_t)(u64A & 0xffffffff) * u32B;
430 uint64_t u64Hi = (uint64_t)(u64A >> 32) * u32B;
431 u64Hi += (u64Lo >> 32);
432 u.s.Hi = (uint32_t)(u64Hi / u32C);
433 u.s.Lo = (uint32_t)((((u64Hi % u32C) << 32) + (u64Lo & 0xffffffff)) / u32C);
434 return u.u;
435# endif
436}
437#endif
438
439/** @} */
440#endif /* !IPRT_INCLUDED_asm_math_h */
441
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