• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright 2015, VIXL authors
2 // All rights reserved.
3 //
4 // Redistribution and use in source and binary forms, with or without
5 // modification, are permitted provided that the following conditions are met:
6 //
7 //   * Redistributions of source code must retain the above copyright notice,
8 //     this list of conditions and the following disclaimer.
9 //   * Redistributions in binary form must reproduce the above copyright notice,
10 //     this list of conditions and the following disclaimer in the documentation
11 //     and/or other materials provided with the distribution.
12 //   * Neither the name of ARM Limited nor the names of its contributors may be
13 //     used to endorse or promote products derived from this software without
14 //     specific prior written permission.
15 //
16 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS CONTRIBUTORS "AS IS" AND
17 // ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
18 // WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
19 // DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE
20 // FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 // DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
22 // SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
23 // CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
24 // OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
25 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26 
27 #if defined(__aarch64__) && (defined(__ANDROID__) || defined(__linux__))
28 #include <sys/auxv.h>
29 #define VIXL_USE_LINUX_HWCAP 1
30 #endif
31 
32 #include "../utils-vixl.h"
33 
34 #include "cpu-aarch64.h"
35 
36 namespace vixl {
37 namespace aarch64 {
38 
39 
40 const IDRegister::Field AA64PFR0::kFP(16, Field::kSigned);
41 const IDRegister::Field AA64PFR0::kAdvSIMD(20, Field::kSigned);
42 const IDRegister::Field AA64PFR0::kRAS(28);
43 const IDRegister::Field AA64PFR0::kSVE(32);
44 const IDRegister::Field AA64PFR0::kDIT(48);
45 const IDRegister::Field AA64PFR0::kCSV2(56);
46 const IDRegister::Field AA64PFR0::kCSV3(60);
47 
48 const IDRegister::Field AA64PFR1::kBT(0);
49 const IDRegister::Field AA64PFR1::kSSBS(4);
50 const IDRegister::Field AA64PFR1::kMTE(8);
51 
52 const IDRegister::Field AA64ISAR0::kAES(4);
53 const IDRegister::Field AA64ISAR0::kSHA1(8);
54 const IDRegister::Field AA64ISAR0::kSHA2(12);
55 const IDRegister::Field AA64ISAR0::kCRC32(16);
56 const IDRegister::Field AA64ISAR0::kAtomic(20);
57 const IDRegister::Field AA64ISAR0::kRDM(28);
58 const IDRegister::Field AA64ISAR0::kSHA3(32);
59 const IDRegister::Field AA64ISAR0::kSM3(36);
60 const IDRegister::Field AA64ISAR0::kSM4(40);
61 const IDRegister::Field AA64ISAR0::kDP(44);
62 const IDRegister::Field AA64ISAR0::kFHM(48);
63 const IDRegister::Field AA64ISAR0::kTS(52);
64 const IDRegister::Field AA64ISAR0::kRNDR(60);
65 
66 const IDRegister::Field AA64ISAR1::kDPB(0);
67 const IDRegister::Field AA64ISAR1::kAPA(4);
68 const IDRegister::Field AA64ISAR1::kAPI(8);
69 const IDRegister::Field AA64ISAR1::kJSCVT(12);
70 const IDRegister::Field AA64ISAR1::kFCMA(16);
71 const IDRegister::Field AA64ISAR1::kLRCPC(20);
72 const IDRegister::Field AA64ISAR1::kGPA(24);
73 const IDRegister::Field AA64ISAR1::kGPI(28);
74 const IDRegister::Field AA64ISAR1::kFRINTTS(32);
75 const IDRegister::Field AA64ISAR1::kSB(36);
76 const IDRegister::Field AA64ISAR1::kSPECRES(40);
77 const IDRegister::Field AA64ISAR1::kBF16(44);
78 const IDRegister::Field AA64ISAR1::kDGH(48);
79 const IDRegister::Field AA64ISAR1::kI8MM(52);
80 
81 const IDRegister::Field AA64ISAR2::kRPRES(4);
82 
83 const IDRegister::Field AA64MMFR0::kECV(60);
84 
85 const IDRegister::Field AA64MMFR1::kLO(16);
86 const IDRegister::Field AA64MMFR1::kAFP(44);
87 
88 const IDRegister::Field AA64MMFR2::kAT(32);
89 
90 const IDRegister::Field AA64ZFR0::kSVEver(0);
91 const IDRegister::Field AA64ZFR0::kAES(4);
92 const IDRegister::Field AA64ZFR0::kBitPerm(16);
93 const IDRegister::Field AA64ZFR0::kBF16(20);
94 const IDRegister::Field AA64ZFR0::kSHA3(32);
95 const IDRegister::Field AA64ZFR0::kSM4(40);
96 const IDRegister::Field AA64ZFR0::kI8MM(44);
97 const IDRegister::Field AA64ZFR0::kF32MM(52);
98 const IDRegister::Field AA64ZFR0::kF64MM(56);
99 
GetCPUFeatures() const100 CPUFeatures AA64PFR0::GetCPUFeatures() const {
101   CPUFeatures f;
102   if (Get(kFP) >= 0) f.Combine(CPUFeatures::kFP);
103   if (Get(kFP) >= 1) f.Combine(CPUFeatures::kFPHalf);
104   if (Get(kAdvSIMD) >= 0) f.Combine(CPUFeatures::kNEON);
105   if (Get(kAdvSIMD) >= 1) f.Combine(CPUFeatures::kNEONHalf);
106   if (Get(kRAS) >= 1) f.Combine(CPUFeatures::kRAS);
107   if (Get(kSVE) >= 1) f.Combine(CPUFeatures::kSVE);
108   if (Get(kDIT) >= 1) f.Combine(CPUFeatures::kDIT);
109   if (Get(kCSV2) >= 1) f.Combine(CPUFeatures::kCSV2);
110   if (Get(kCSV2) >= 2) f.Combine(CPUFeatures::kSCXTNUM);
111   if (Get(kCSV3) >= 1) f.Combine(CPUFeatures::kCSV3);
112   return f;
113 }
114 
GetCPUFeatures() const115 CPUFeatures AA64PFR1::GetCPUFeatures() const {
116   CPUFeatures f;
117   if (Get(kBT) >= 1) f.Combine(CPUFeatures::kBTI);
118   if (Get(kSSBS) >= 1) f.Combine(CPUFeatures::kSSBS);
119   if (Get(kSSBS) >= 2) f.Combine(CPUFeatures::kSSBSControl);
120   if (Get(kMTE) >= 1) f.Combine(CPUFeatures::kMTEInstructions);
121   if (Get(kMTE) >= 2) f.Combine(CPUFeatures::kMTE);
122   return f;
123 }
124 
GetCPUFeatures() const125 CPUFeatures AA64ISAR0::GetCPUFeatures() const {
126   CPUFeatures f;
127   if (Get(kAES) >= 1) f.Combine(CPUFeatures::kAES);
128   if (Get(kAES) >= 2) f.Combine(CPUFeatures::kPmull1Q);
129   if (Get(kSHA1) >= 1) f.Combine(CPUFeatures::kSHA1);
130   if (Get(kSHA2) >= 1) f.Combine(CPUFeatures::kSHA2);
131   if (Get(kSHA2) >= 2) f.Combine(CPUFeatures::kSHA512);
132   if (Get(kCRC32) >= 1) f.Combine(CPUFeatures::kCRC32);
133   if (Get(kAtomic) >= 1) f.Combine(CPUFeatures::kAtomics);
134   if (Get(kRDM) >= 1) f.Combine(CPUFeatures::kRDM);
135   if (Get(kSHA3) >= 1) f.Combine(CPUFeatures::kSHA3);
136   if (Get(kSM3) >= 1) f.Combine(CPUFeatures::kSM3);
137   if (Get(kSM4) >= 1) f.Combine(CPUFeatures::kSM4);
138   if (Get(kDP) >= 1) f.Combine(CPUFeatures::kDotProduct);
139   if (Get(kFHM) >= 1) f.Combine(CPUFeatures::kFHM);
140   if (Get(kTS) >= 1) f.Combine(CPUFeatures::kFlagM);
141   if (Get(kTS) >= 2) f.Combine(CPUFeatures::kAXFlag);
142   if (Get(kRNDR) >= 1) f.Combine(CPUFeatures::kRNG);
143   return f;
144 }
145 
GetCPUFeatures() const146 CPUFeatures AA64ISAR1::GetCPUFeatures() const {
147   CPUFeatures f;
148   if (Get(kDPB) >= 1) f.Combine(CPUFeatures::kDCPoP);
149   if (Get(kDPB) >= 2) f.Combine(CPUFeatures::kDCCVADP);
150   if (Get(kJSCVT) >= 1) f.Combine(CPUFeatures::kJSCVT);
151   if (Get(kFCMA) >= 1) f.Combine(CPUFeatures::kFcma);
152   if (Get(kLRCPC) >= 1) f.Combine(CPUFeatures::kRCpc);
153   if (Get(kLRCPC) >= 2) f.Combine(CPUFeatures::kRCpcImm);
154   if (Get(kFRINTTS) >= 1) f.Combine(CPUFeatures::kFrintToFixedSizedInt);
155   if (Get(kSB) >= 1) f.Combine(CPUFeatures::kSB);
156   if (Get(kSPECRES) >= 1) f.Combine(CPUFeatures::kSPECRES);
157   if (Get(kBF16) >= 1) f.Combine(CPUFeatures::kBF16);
158   if (Get(kDGH) >= 1) f.Combine(CPUFeatures::kDGH);
159   if (Get(kI8MM) >= 1) f.Combine(CPUFeatures::kI8MM);
160 
161   // Only one of these fields should be non-zero, but they have the same
162   // encodings, so merge the logic.
163   int apx = std::max(Get(kAPI), Get(kAPA));
164   if (apx >= 1) {
165     f.Combine(CPUFeatures::kPAuth);
166     // APA (rather than API) indicates QARMA.
167     if (Get(kAPA) >= 1) f.Combine(CPUFeatures::kPAuthQARMA);
168     if (apx == 0b0010) f.Combine(CPUFeatures::kPAuthEnhancedPAC);
169     if (apx >= 0b0011) f.Combine(CPUFeatures::kPAuthEnhancedPAC2);
170     if (apx >= 0b0100) f.Combine(CPUFeatures::kPAuthFPAC);
171     if (apx >= 0b0101) f.Combine(CPUFeatures::kPAuthFPACCombined);
172   }
173 
174   if (Get(kGPI) >= 1) f.Combine(CPUFeatures::kPAuthGeneric);
175   if (Get(kGPA) >= 1) {
176     f.Combine(CPUFeatures::kPAuthGeneric, CPUFeatures::kPAuthGenericQARMA);
177   }
178   return f;
179 }
180 
GetCPUFeatures() const181 CPUFeatures AA64ISAR2::GetCPUFeatures() const {
182   CPUFeatures f;
183   if (Get(kRPRES) >= 1) f.Combine(CPUFeatures::kRPRES);
184   return f;
185 }
186 
GetCPUFeatures() const187 CPUFeatures AA64MMFR0::GetCPUFeatures() const {
188   CPUFeatures f;
189   if (Get(kECV) >= 1) f.Combine(CPUFeatures::kECV);
190   return f;
191 }
192 
GetCPUFeatures() const193 CPUFeatures AA64MMFR1::GetCPUFeatures() const {
194   CPUFeatures f;
195   if (Get(kLO) >= 1) f.Combine(CPUFeatures::kLORegions);
196   if (Get(kAFP) >= 1) f.Combine(CPUFeatures::kAFP);
197   return f;
198 }
199 
GetCPUFeatures() const200 CPUFeatures AA64MMFR2::GetCPUFeatures() const {
201   CPUFeatures f;
202   if (Get(kAT) >= 1) f.Combine(CPUFeatures::kUSCAT);
203   return f;
204 }
205 
GetCPUFeatures() const206 CPUFeatures AA64ZFR0::GetCPUFeatures() const {
207   // This register is only available with SVE, but reads-as-zero in its absence,
208   // so it's always safe to read it.
209   CPUFeatures f;
210   if (Get(kF64MM) >= 1) f.Combine(CPUFeatures::kSVEF64MM);
211   if (Get(kF32MM) >= 1) f.Combine(CPUFeatures::kSVEF32MM);
212   if (Get(kI8MM) >= 1) f.Combine(CPUFeatures::kSVEI8MM);
213   if (Get(kSM4) >= 1) f.Combine(CPUFeatures::kSVESM4);
214   if (Get(kSHA3) >= 1) f.Combine(CPUFeatures::kSVESHA3);
215   if (Get(kBF16) >= 1) f.Combine(CPUFeatures::kSVEBF16);
216   if (Get(kBitPerm) >= 1) f.Combine(CPUFeatures::kSVEBitPerm);
217   if (Get(kAES) >= 1) f.Combine(CPUFeatures::kSVEAES);
218   if (Get(kAES) >= 2) f.Combine(CPUFeatures::kSVEPmull128);
219   if (Get(kSVEver) >= 1) f.Combine(CPUFeatures::kSVE2);
220   return f;
221 }
222 
Get(IDRegister::Field field) const223 int IDRegister::Get(IDRegister::Field field) const {
224   int msb = field.GetMsb();
225   int lsb = field.GetLsb();
226   VIXL_STATIC_ASSERT(static_cast<size_t>(Field::kMaxWidthInBits) <
227                      (sizeof(int) * kBitsPerByte));
228   switch (field.GetType()) {
229     case Field::kSigned:
230       return static_cast<int>(ExtractSignedBitfield64(msb, lsb, value_));
231     case Field::kUnsigned:
232       return static_cast<int>(ExtractUnsignedBitfield64(msb, lsb, value_));
233   }
234   VIXL_UNREACHABLE();
235   return 0;
236 }
237 
InferCPUFeaturesFromIDRegisters()238 CPUFeatures CPU::InferCPUFeaturesFromIDRegisters() {
239   CPUFeatures f;
240 #define VIXL_COMBINE_ID_REG(NAME, MRS_ARG) \
241   f.Combine(Read##NAME().GetCPUFeatures());
242   VIXL_AARCH64_ID_REG_LIST(VIXL_COMBINE_ID_REG)
243 #undef VIXL_COMBINE_ID_REG
244   return f;
245 }
246 
InferCPUFeaturesFromOS(CPUFeatures::QueryIDRegistersOption option)247 CPUFeatures CPU::InferCPUFeaturesFromOS(
248     CPUFeatures::QueryIDRegistersOption option) {
249   CPUFeatures features;
250 
251 #if VIXL_USE_LINUX_HWCAP
252   // Map each set bit onto a feature. Ideally, we'd use HWCAP_* macros rather
253   // than explicit bits, but explicit bits allow us to identify features that
254   // the toolchain doesn't know about.
255   static const CPUFeatures::Feature kFeatureBits[] =
256       {// Bits 0-7
257        CPUFeatures::kFP,
258        CPUFeatures::kNEON,
259        CPUFeatures::kNone,  // "EVTSTRM", which VIXL doesn't track.
260        CPUFeatures::kAES,
261        CPUFeatures::kPmull1Q,
262        CPUFeatures::kSHA1,
263        CPUFeatures::kSHA2,
264        CPUFeatures::kCRC32,
265        // Bits 8-15
266        CPUFeatures::kAtomics,
267        CPUFeatures::kFPHalf,
268        CPUFeatures::kNEONHalf,
269        CPUFeatures::kIDRegisterEmulation,
270        CPUFeatures::kRDM,
271        CPUFeatures::kJSCVT,
272        CPUFeatures::kFcma,
273        CPUFeatures::kRCpc,
274        // Bits 16-23
275        CPUFeatures::kDCPoP,
276        CPUFeatures::kSHA3,
277        CPUFeatures::kSM3,
278        CPUFeatures::kSM4,
279        CPUFeatures::kDotProduct,
280        CPUFeatures::kSHA512,
281        CPUFeatures::kSVE,
282        CPUFeatures::kFHM,
283        // Bits 24-31
284        CPUFeatures::kDIT,
285        CPUFeatures::kUSCAT,
286        CPUFeatures::kRCpcImm,
287        CPUFeatures::kFlagM,
288        CPUFeatures::kSSBSControl,
289        CPUFeatures::kSB,
290        CPUFeatures::kPAuth,
291        CPUFeatures::kPAuthGeneric,
292        // Bits 32-39
293        CPUFeatures::kDCCVADP,
294        CPUFeatures::kSVE2,
295        CPUFeatures::kSVEAES,
296        CPUFeatures::kSVEPmull128,
297        CPUFeatures::kSVEBitPerm,
298        CPUFeatures::kSVESHA3,
299        CPUFeatures::kSVESM4,
300        CPUFeatures::kAXFlag,
301        // Bits 40-47
302        CPUFeatures::kFrintToFixedSizedInt,
303        CPUFeatures::kSVEI8MM,
304        CPUFeatures::kSVEF32MM,
305        CPUFeatures::kSVEF64MM,
306        CPUFeatures::kSVEBF16,
307        CPUFeatures::kI8MM,
308        CPUFeatures::kBF16,
309        CPUFeatures::kDGH,
310        // Bits 48+
311        CPUFeatures::kRNG,
312        CPUFeatures::kBTI,
313        CPUFeatures::kMTE,
314        CPUFeatures::kECV,
315        CPUFeatures::kAFP,
316        CPUFeatures::kRPRES};
317 
318   uint64_t hwcap_low32 = getauxval(AT_HWCAP);
319   uint64_t hwcap_high32 = getauxval(AT_HWCAP2);
320   VIXL_ASSERT(IsUint32(hwcap_low32));
321   VIXL_ASSERT(IsUint32(hwcap_high32));
322   uint64_t hwcap = hwcap_low32 | (hwcap_high32 << 32);
323 
324   VIXL_STATIC_ASSERT(ArrayLength(kFeatureBits) < 64);
325   for (size_t i = 0; i < ArrayLength(kFeatureBits); i++) {
326     if (hwcap & (UINT64_C(1) << i)) features.Combine(kFeatureBits[i]);
327   }
328   // MTE support from HWCAP2 signifies FEAT_MTE1 and FEAT_MTE2 support
329   if (features.Has(CPUFeatures::kMTE)) {
330     features.Combine(CPUFeatures::kMTEInstructions);
331   }
332 #endif  // VIXL_USE_LINUX_HWCAP
333 
334   if ((option == CPUFeatures::kQueryIDRegistersIfAvailable) &&
335       (features.Has(CPUFeatures::kIDRegisterEmulation))) {
336     features.Combine(InferCPUFeaturesFromIDRegisters());
337   }
338   return features;
339 }
340 
341 
342 #ifdef __aarch64__
343 #define VIXL_READ_ID_REG(NAME, MRS_ARG)        \
344   NAME CPU::Read##NAME() {                     \
345     uint64_t value = 0;                        \
346     __asm__("mrs %0, " MRS_ARG : "=r"(value)); \
347     return NAME(value);                        \
348   }
349 #else  // __aarch64__
350 #define VIXL_READ_ID_REG(NAME, MRS_ARG) \
351   NAME CPU::Read##NAME() {              \
352     VIXL_UNREACHABLE();                 \
353     return NAME(0);                     \
354   }
355 #endif  // __aarch64__
356 
357 VIXL_AARCH64_ID_REG_LIST(VIXL_READ_ID_REG)
358 
359 #undef VIXL_READ_ID_REG
360 
361 
362 // Initialise to smallest possible cache size.
363 unsigned CPU::dcache_line_size_ = 1;
364 unsigned CPU::icache_line_size_ = 1;
365 
366 
367 // Currently computes I and D cache line size.
SetUp()368 void CPU::SetUp() {
369   uint32_t cache_type_register = GetCacheType();
370 
371   // The cache type register holds information about the caches, including I
372   // D caches line size.
373   static const int kDCacheLineSizeShift = 16;
374   static const int kICacheLineSizeShift = 0;
375   static const uint32_t kDCacheLineSizeMask = 0xf << kDCacheLineSizeShift;
376   static const uint32_t kICacheLineSizeMask = 0xf << kICacheLineSizeShift;
377 
378   // The cache type register holds the size of the I and D caches in words as
379   // a power of two.
380   uint32_t dcache_line_size_power_of_two =
381       (cache_type_register & kDCacheLineSizeMask) >> kDCacheLineSizeShift;
382   uint32_t icache_line_size_power_of_two =
383       (cache_type_register & kICacheLineSizeMask) >> kICacheLineSizeShift;
384 
385   dcache_line_size_ = 4 << dcache_line_size_power_of_two;
386   icache_line_size_ = 4 << icache_line_size_power_of_two;
387 }
388 
389 
GetCacheType()390 uint32_t CPU::GetCacheType() {
391 #ifdef __aarch64__
392   uint64_t cache_type_register;
393   // Copy the content of the cache type register to a core register.
394   __asm__ __volatile__("mrs %[ctr], ctr_el0"  // NOLINT(runtime/references)
395                        : [ctr] "=r"(cache_type_register));
396   VIXL_ASSERT(IsUint32(cache_type_register));
397   return static_cast<uint32_t>(cache_type_register);
398 #else
399   // This will lead to a cache with 1 byte long lines, which is fine since
400   // neither EnsureIAndDCacheCoherency nor the simulator will need this
401   // information.
402   return 0;
403 #endif
404 }
405 
406 
407 // Query the SVE vector length. This requires CPUFeatures::kSVE.
ReadSVEVectorLengthInBits()408 int CPU::ReadSVEVectorLengthInBits() {
409 #ifdef __aarch64__
410   uint64_t vl;
411   // To support compilers that don't understand `rdvl`, encode the value
412   // directly and move it manually.
413   __asm__(
414       "   .word 0x04bf5100\n"  // rdvl x0, #8
415       "   mov %[vl], x0\n"
416       : [vl] "=r"(vl)
417       :
418       : "x0");
419   VIXL_ASSERT(vl <= INT_MAX);
420   return static_cast<int>(vl);
421 #else
422   VIXL_UNREACHABLE();
423   return 0;
424 #endif
425 }
426 
427 
EnsureIAndDCacheCoherency(void * address,size_t length)428 void CPU::EnsureIAndDCacheCoherency(void *address, size_t length) {
429 #ifdef __aarch64__
430   // Implement the cache synchronisation for all targets where AArch64 is the
431   // host, even if we're building the simulator for an AAarch64 host. This
432   // allows for cases where the user wants to simulate code as well as run it
433   // natively.
434 
435   if (length == 0) {
436     return;
437   }
438 
439   // The code below assumes user space cache operations are allowed.
440 
441   // Work out the line sizes for each cache, and use them to determine the
442   // start addresses.
443   uintptr_t start = reinterpret_cast<uintptr_t>(address);
444   uintptr_t dsize = static_cast<uintptr_t>(dcache_line_size_);
445   uintptr_t isize = static_cast<uintptr_t>(icache_line_size_);
446   uintptr_t dline = start & ~(dsize - 1);
447   uintptr_t iline = start & ~(isize - 1);
448 
449   // Cache line sizes are always a power of 2.
450   VIXL_ASSERT(IsPowerOf2(dsize));
451   VIXL_ASSERT(IsPowerOf2(isize));
452   uintptr_t end = start + length;
453 
454   do {
455     __asm__ __volatile__(
456         // Clean each line of the D cache containing the target data.
457         //
458         // dc       : Data Cache maintenance
459         //     c    : Clean
460         //      va  : by (Virtual) Address
461         //        u : to the point of Unification
462         // The point of unification for a processor is the point by which the
463         // instruction and data caches are guaranteed to see the same copy of a
464         // memory location. See ARM DDI 0406B page B2-12 for more information.
465         "   dc    cvau, %[dline]\n"
466         :
467         : [dline] "r"(dline)
468         // This code does not write to memory, but the "memory" dependency
469         // prevents GCC from reordering the code.
470         : "memory");
471     dline += dsize;
472   } while (dline < end);
473 
474   __asm__ __volatile__(
475       // Make sure that the data cache operations (above) complete before the
476       // instruction cache operations (below).
477       //
478       // dsb      : Data Synchronisation Barrier
479       //      ish : Inner SHareable domain
480       //
481       // The point of unification for an Inner Shareable shareability domain is
482       // the point by which the instruction and data caches of all the
483       // processors
484       // in that Inner Shareable shareability domain are guaranteed to see the
485       // same copy of a memory location. See ARM DDI 0406B page B2-12 for more
486       // information.
487       "   dsb   ish\n"
488       :
489       :
490       : "memory");
491 
492   do {
493     __asm__ __volatile__(
494         // Invalidate each line of the I cache containing the target data.
495         //
496         // ic      : Instruction Cache maintenance
497         //    i    : Invalidate
498         //     va  : by Address
499         //       u : to the point of Unification
500         "   ic   ivau, %[iline]\n"
501         :
502         : [iline] "r"(iline)
503         : "memory");
504     iline += isize;
505   } while (iline < end);
506 
507   __asm__ __volatile__(
508       // Make sure that the instruction cache operations (above) take effect
509       // before the isb (below).
510       "   dsb  ish\n"
511 
512       // Ensure that any instructions already in the pipeline are discarded and
513       // reloaded from the new data.
514       // isb : Instruction Synchronisation Barrier
515       "   isb\n"
516       :
517       :
518       : "memory");
519 #else
520   // If the host isn't AArch64, we must be using the simulator, so this function
521   // doesn't have to do anything.
522   USE(address, length);
523 #endif
524 }
525 
526 }  // namespace aarch64
527 }  // namespace vixl
528