VirtualBox

source: vbox/trunk/include/iprt/asm-math.h@ 83776

Last change on this file since 83776 was 83776, checked in by vboxsync, 4 years ago

iprt/asm*,++: Contain the intrin.h mess in iprt/win/msvc_intrin.h. Hack it for bs3kit and no-crt to avoid dragging in malloc.h from UCRT with VC++ 14.1. bugref:8489
  • Property svn:eol-style set to native
  • Property svn:keywords set to Author Date Id Revision
File size: 13.1 KB
Line 
1/** @file
2 * IPRT - Assembly Routines for Optimizing some Integers Math Operations.
3 */
4
5/*
6 * Copyright (C) 2006-2020 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/* Emit the intrinsics at all optimization levels. */
36# include <iprt/win/msvc_intrin.h>
37# pragma intrinsic(__emul)
38# pragma intrinsic(__emulu)
39# ifdef RT_ARCH_AMD64
40# pragma intrinsic(_mul128)
41# pragma intrinsic(_umul128)
42# endif
43#endif
44
45
46/** @defgroup grp_rt_asm_math Interger Math Optimizations
47 * @ingroup grp_rt_asm
48 * @{ */
49
50/**
51 * Multiplies two unsigned 32-bit values returning an unsigned 64-bit result.
52 *
53 * @returns u32F1 * u32F2.
54 */
55
56#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_X86)
57DECLASM(uint64_t) ASMMult2xU32RetU64(uint32_t u32F1, uint32_t u32F2);
58#else
59DECLINLINE(uint64_t) ASMMult2xU32RetU64(uint32_t u32F1, uint32_t u32F2)
60{
61# ifdef RT_ARCH_X86
62 uint64_t u64;
63# if RT_INLINE_ASM_GNU_STYLE
64 __asm__ __volatile__("mull %%edx"
65 : "=A" (u64)
66 : "a" (u32F2), "d" (u32F1));
67# elif RT_INLINE_ASM_USES_INTRIN
68 u64 = __emulu(u32F1, u32F2);
69# else
70 __asm
71 {
72 mov edx, [u32F1]
73 mov eax, [u32F2]
74 mul edx
75 mov dword ptr [u64], eax
76 mov dword ptr [u64 + 4], edx
77 }
78# endif
79 return u64;
80# else /* generic: */
81 return (uint64_t)u32F1 * u32F2;
82# endif
83}
84#endif
85
86
87/**
88 * Multiplies two signed 32-bit values returning a signed 64-bit result.
89 *
90 * @returns u32F1 * u32F2.
91 */
92#if RT_INLINE_ASM_EXTERNAL && !RT_INLINE_ASM_USES_INTRIN && defined(RT_ARCH_X86)
93DECLASM(int64_t) ASMMult2xS32RetS64(int32_t i32F1, int32_t i32F2);
94#else
95DECLINLINE(int64_t) ASMMult2xS32RetS64(int32_t i32F1, int32_t i32F2)
96{
97# ifdef RT_ARCH_X86
98 int64_t i64;
99# if RT_INLINE_ASM_GNU_STYLE
100 __asm__ __volatile__("imull %%edx"
101 : "=A" (i64)
102 : "a" (i32F2), "d" (i32F1));
103# elif RT_INLINE_ASM_USES_INTRIN
104 i64 = __emul(i32F1, i32F2);
105# else
106 __asm
107 {
108 mov edx, [i32F1]
109 mov eax, [i32F2]
110 imul edx
111 mov dword ptr [i64], eax
112 mov dword ptr [i64 + 4], edx
113 }
114# endif
115 return i64;
116# else /* generic: */
117 return (int64_t)i32F1 * i32F2;
118# endif
119}
120#endif
121
122
123#if ARCH_BITS == 64
124DECLINLINE(uint64_t) ASMMult2xU64Ret2xU64(uint64_t u64F1, uint64_t u64F2, uint64_t *pu64ProdHi)
125{
126# if defined(RT_ARCH_AMD64) && (RT_INLINE_ASM_GNU_STYLE || RT_INLINE_ASM_USES_INTRIN)
127# if RT_INLINE_ASM_GNU_STYLE
128 uint64_t u64Low, u64High;
129 __asm__ __volatile__("mulq %%rdx"
130 : "=a" (u64Low), "=d" (u64High)
131 : "0" (u64F1), "1" (u64F2));
132 *pu64ProdHi = u64High;
133 return u64Low;
134# elif RT_INLINE_ASM_USES_INTRIN
135 return _umul128(u64F1, u64F2, pu64ProdHi);
136# else
137# error "hmm"
138# endif
139# else /* generic: */
140 /*
141 * F1 * F2 = Prod
142 * -- --
143 * ab * cd = b*d + a*d*10 + b*c*10 + a*c*100
144 *
145 * Where a, b, c and d are 'digits', and 10 is max digit + 1.
146 *
147 * Our digits are 32-bit wide, so instead of 10 we multiply by 4G.
148 * Prod = F1.s.Lo*F2.s.Lo + F1.s.Hi*F2.s.Lo*4G
149 * + F1.s.Lo*F2.s.Hi*4G + F1.s.Hi*F2.s.Hi*4G*4G
150 */
151 RTUINT128U Prod;
152 RTUINT64U Tmp1;
153 uint64_t u64Tmp;
154 RTUINT64U F1, F2;
155 F1.u = u64F1;
156 F2.u = u64F2;
157
158 Prod.s.Lo = ASMMult2xU32RetU64(F1.s.Lo, F2.s.Lo);
159
160 Tmp1.u = ASMMult2xU32RetU64(F1.s.Hi, F2.s.Lo);
161 u64Tmp = (uint64_t)Prod.DWords.dw1 + Tmp1.s.Lo;
162 Prod.DWords.dw1 = (uint32_t)u64Tmp;
163 Prod.s.Hi = Tmp1.s.Hi;
164 Prod.s.Hi += u64Tmp >> 32; /* carry */
165
166 Tmp1.u = ASMMult2xU32RetU64(F1.s.Lo, F2.s.Hi);
167 u64Tmp = (uint64_t)Prod.DWords.dw1 + Tmp1.s.Lo;
168 Prod.DWords.dw1 = (uint32_t)u64Tmp;
169 u64Tmp >>= 32; /* carry */
170 u64Tmp += Prod.DWords.dw2;
171 u64Tmp += Tmp1.s.Hi;
172 Prod.DWords.dw2 = (uint32_t)u64Tmp;
173 Prod.DWords.dw3 += u64Tmp >> 32; /* carry */
174
175 Prod.s.Hi += ASMMult2xU32RetU64(F1.s.Hi, F2.s.Hi);
176 *pu64ProdHi = Prod.s.Hi;
177 return Prod.s.Lo;
178# endif
179}
180#endif
181
182
183
184/**
185 * Divides a 64-bit unsigned by a 32-bit unsigned returning an unsigned 32-bit result.
186 *
187 * @returns u64 / u32.
188 */
189#if RT_INLINE_ASM_EXTERNAL && defined(RT_ARCH_X86)
190DECLASM(uint32_t) ASMDivU64ByU32RetU32(uint64_t u64, uint32_t u32);
191#else
192DECLINLINE(uint32_t) ASMDivU64ByU32RetU32(uint64_t u64, uint32_t u32)
193{
194# ifdef RT_ARCH_X86
195# if RT_INLINE_ASM_GNU_STYLE
196 RTCCUINTREG uDummy;
197 __asm__ __volatile__("divl %3"
198 : "=a" (u32), "=d"(uDummy)
199 : "A" (u64), "r" (u32));
200# else
201 __asm
202 {
203 mov eax, dword ptr [u64]
204 mov edx, dword ptr [u64 + 4]
205 mov ecx, [u32]
206 div ecx
207 mov [u32], eax
208 }
209# endif
210 return u32;
211# else /* generic: */
212 return (uint32_t)(u64 / u32);
213# endif
214}
215#endif
216
217
218/**
219 * Divides a 64-bit signed by a 32-bit signed returning a signed 32-bit result.
220 *
221 * @returns u64 / u32.
222 */
223#if RT_INLINE_ASM_EXTERNAL && defined(RT_ARCH_X86)
224DECLASM(int32_t) ASMDivS64ByS32RetS32(int64_t i64, int32_t i32);
225#else
226DECLINLINE(int32_t) ASMDivS64ByS32RetS32(int64_t i64, int32_t i32)
227{
228# ifdef RT_ARCH_X86
229# if RT_INLINE_ASM_GNU_STYLE
230 RTCCUINTREG iDummy;
231 __asm__ __volatile__("idivl %3"
232 : "=a" (i32), "=d"(iDummy)
233 : "A" (i64), "r" (i32));
234# else
235 __asm
236 {
237 mov eax, dword ptr [i64]
238 mov edx, dword ptr [i64 + 4]
239 mov ecx, [i32]
240 idiv ecx
241 mov [i32], eax
242 }
243# endif
244 return i32;
245# else /* generic: */
246 return (int32_t)(i64 / i32);
247# endif
248}
249#endif
250
251
252/**
253 * Performs 64-bit unsigned by a 32-bit unsigned division with a 32-bit unsigned result,
254 * returning the rest.
255 *
256 * @returns u64 % u32.
257 *
258 * @remarks It is important that the result is <= UINT32_MAX or we'll overflow and crash.
259 */
260#if RT_INLINE_ASM_EXTERNAL && defined(RT_ARCH_X86)
261DECLASM(uint32_t) ASMModU64ByU32RetU32(uint64_t u64, uint32_t u32);
262#else
263DECLINLINE(uint32_t) ASMModU64ByU32RetU32(uint64_t u64, uint32_t u32)
264{
265# ifdef RT_ARCH_X86
266# if RT_INLINE_ASM_GNU_STYLE
267 RTCCUINTREG uDummy;
268 __asm__ __volatile__("divl %3"
269 : "=a" (uDummy), "=d"(u32)
270 : "A" (u64), "r" (u32));
271# else
272 __asm
273 {
274 mov eax, dword ptr [u64]
275 mov edx, dword ptr [u64 + 4]
276 mov ecx, [u32]
277 div ecx
278 mov [u32], edx
279 }
280# endif
281 return u32;
282# else /* generic: */
283 return (uint32_t)(u64 % u32);
284# endif
285}
286#endif
287
288
289/**
290 * Performs 64-bit signed by a 32-bit signed division with a 32-bit signed result,
291 * returning the rest.
292 *
293 * @returns u64 % u32.
294 *
295 * @remarks It is important that the result is <= UINT32_MAX or we'll overflow and crash.
296 */
297#if RT_INLINE_ASM_EXTERNAL && defined(RT_ARCH_X86)
298DECLASM(int32_t) ASMModS64ByS32RetS32(int64_t i64, int32_t i32);
299#else
300DECLINLINE(int32_t) ASMModS64ByS32RetS32(int64_t i64, int32_t i32)
301{
302# ifdef RT_ARCH_X86
303# if RT_INLINE_ASM_GNU_STYLE
304 RTCCUINTREG iDummy;
305 __asm__ __volatile__("idivl %3"
306 : "=a" (iDummy), "=d"(i32)
307 : "A" (i64), "r" (i32));
308# else
309 __asm
310 {
311 mov eax, dword ptr [i64]
312 mov edx, dword ptr [i64 + 4]
313 mov ecx, [i32]
314 idiv ecx
315 mov [i32], edx
316 }
317# endif
318 return i32;
319# else /* generic: */
320 return (int32_t)(i64 % i32);
321# endif
322}
323#endif
324
325
326/**
327 * Multiple a 32-bit by a 32-bit integer and divide the result by a 32-bit integer
328 * using a 64 bit intermediate result.
329 *
330 * @returns (u32A * u32B) / u32C.
331 * @param u32A The 32-bit value (A).
332 * @param u32B The 32-bit value to multiple by A.
333 * @param u32C The 32-bit value to divide A*B by.
334 *
335 * @remarks Architecture specific.
336 * @remarks Make sure the result won't ever exceed 32-bit, because hardware
337 * exception may be raised if it does.
338 * @remarks On x86 this may be used to avoid dragging in 64-bit builtin
339 * arithmetics functions.
340 */
341#if RT_INLINE_ASM_EXTERNAL && (defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86))
342DECLASM(uint32_t) ASMMultU32ByU32DivByU32(uint32_t u32A, uint32_t u32B, uint32_t u32C);
343#else
344DECLINLINE(uint32_t) ASMMultU32ByU32DivByU32(uint32_t u32A, uint32_t u32B, uint32_t u32C)
345{
346# if RT_INLINE_ASM_GNU_STYLE && (defined(RT_ARCH_AMD64) || defined(RT_ARCH_X86))
347 uint32_t u32Result, u32Spill;
348 __asm__ __volatile__("mull %2\n\t"
349 "divl %3\n\t"
350 : "=&a" (u32Result),
351 "=&d" (u32Spill)
352 : "r" (u32B),
353 "r" (u32C),
354 "0" (u32A));
355 return u32Result;
356# else
357 return (uint32_t)(((uint64_t)u32A * u32B) / u32C);
358# endif
359}
360#endif
361
362
363/**
364 * Multiple a 64-bit by a 32-bit integer and divide the result by a 32-bit integer
365 * using a 96 bit intermediate result.
366 *
367 * @returns (u64A * u32B) / u32C.
368 * @param u64A The 64-bit value.
369 * @param u32B The 32-bit value to multiple by A.
370 * @param u32C The 32-bit value to divide A*B by.
371 *
372 * @remarks Architecture specific.
373 * @remarks Make sure the result won't ever exceed 64-bit, because hardware
374 * exception may be raised if it does.
375 * @remarks On x86 this may be used to avoid dragging in 64-bit builtin
376 * arithmetics function.
377 */
378#if RT_INLINE_ASM_EXTERNAL || !defined(__GNUC__) || (!defined(RT_ARCH_AMD64) && !defined(RT_ARCH_X86))
379DECLASM(uint64_t) ASMMultU64ByU32DivByU32(uint64_t u64A, uint32_t u32B, uint32_t u32C);
380#else
381DECLINLINE(uint64_t) ASMMultU64ByU32DivByU32(uint64_t u64A, uint32_t u32B, uint32_t u32C)
382{
383# if RT_INLINE_ASM_GNU_STYLE
384# ifdef RT_ARCH_AMD64
385 uint64_t u64Result, u64Spill;
386 __asm__ __volatile__("mulq %2\n\t"
387 "divq %3\n\t"
388 : "=&a" (u64Result),
389 "=&d" (u64Spill)
390 : "r" ((uint64_t)u32B),
391 "r" ((uint64_t)u32C),
392 "0" (u64A));
393 return u64Result;
394# else
395 uint32_t u32Dummy;
396 uint64_t u64Result;
397 __asm__ __volatile__("mull %%ecx \n\t" /* eax = u64Lo.lo = (u64A.lo * u32B).lo
398 edx = u64Lo.hi = (u64A.lo * u32B).hi */
399 "xchg %%eax,%%esi \n\t" /* esi = u64Lo.lo
400 eax = u64A.hi */
401 "xchg %%edx,%%edi \n\t" /* edi = u64Low.hi
402 edx = u32C */
403 "xchg %%edx,%%ecx \n\t" /* ecx = u32C
404 edx = u32B */
405 "mull %%edx \n\t" /* eax = u64Hi.lo = (u64A.hi * u32B).lo
406 edx = u64Hi.hi = (u64A.hi * u32B).hi */
407 "addl %%edi,%%eax \n\t" /* u64Hi.lo += u64Lo.hi */
408 "adcl $0,%%edx \n\t" /* u64Hi.hi += carry */
409 "divl %%ecx \n\t" /* eax = u64Hi / u32C
410 edx = u64Hi % u32C */
411 "movl %%eax,%%edi \n\t" /* edi = u64Result.hi = u64Hi / u32C */
412 "movl %%esi,%%eax \n\t" /* eax = u64Lo.lo */
413 "divl %%ecx \n\t" /* u64Result.lo */
414 "movl %%edi,%%edx \n\t" /* u64Result.hi */
415 : "=A"(u64Result), "=c"(u32Dummy),
416 "=S"(u32Dummy), "=D"(u32Dummy)
417 : "a"((uint32_t)u64A),
418 "S"((uint32_t)(u64A >> 32)),
419 "c"(u32B),
420 "D"(u32C));
421 return u64Result;
422# endif
423# else
424 RTUINT64U u;
425 uint64_t u64Lo = (uint64_t)(u64A & 0xffffffff) * u32B;
426 uint64_t u64Hi = (uint64_t)(u64A >> 32) * u32B;
427 u64Hi += (u64Lo >> 32);
428 u.s.Hi = (uint32_t)(u64Hi / u32C);
429 u.s.Lo = (uint32_t)((((u64Hi % u32C) << 32) + (u64Lo & 0xffffffff)) / u32C);
430 return u.u;
431# endif
432}
433#endif
434
435/** @} */
436#endif /* !IPRT_INCLUDED_asm_math_h */
437
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