xref: /freebsd/contrib/llvm-project/compiler-rt/lib/builtins/clear_cache.c (revision 094517119c62c23369d545a7475ae982d86330a3)
1 //===-- clear_cache.c - Implement __clear_cache ---------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 
9 #include "int_lib.h"
10 #if defined(__linux__)
11 #include <assert.h>
12 #endif
13 #include <stddef.h>
14 
15 #if __APPLE__
16 #include <libkern/OSCacheControl.h>
17 #endif
18 
19 #if defined(_WIN32)
20 // Forward declare Win32 APIs since the GCC mode driver does not handle the
21 // newer SDKs as well as needed.
22 uint32_t FlushInstructionCache(uintptr_t hProcess, void *lpBaseAddress,
23                                uintptr_t dwSize);
24 uintptr_t GetCurrentProcess(void);
25 #endif
26 
27 #if defined(__FreeBSD__) && defined(__arm__)
28 // clang-format off
29 #include <sys/types.h>
30 #include <machine/sysarch.h>
31 // clang-format on
32 #endif
33 
34 #if defined(__NetBSD__) && defined(__arm__)
35 #include <machine/sysarch.h>
36 #endif
37 
38 #if defined(__OpenBSD__) && (defined(__arm__) || defined(__mips__) || defined(__riscv))
39 // clang-format off
40 #include <sys/types.h>
41 #include <machine/sysarch.h>
42 // clang-format on
43 #endif
44 
45 #if defined(__linux__) && defined(__mips__)
46 #include <sys/cachectl.h>
47 #include <sys/syscall.h>
48 #include <unistd.h>
49 #endif
50 
51 #if defined(__linux__) && defined(__riscv)
52 // to get platform-specific syscall definitions
53 #include <linux/unistd.h>
54 #endif
55 
56 // The compiler generates calls to __clear_cache() when creating
57 // trampoline functions on the stack for use with nested functions.
58 // It is expected to invalidate the instruction cache for the
59 // specified range.
60 
61 void __clear_cache(void *start, void *end) {
62 #if __i386__ || __x86_64__ || defined(_M_IX86) || defined(_M_X64)
63 // Intel processors have a unified instruction and data cache
64 // so there is nothing to do
65 #elif defined(_WIN32) && (defined(__arm__) || defined(__aarch64__))
66   FlushInstructionCache(GetCurrentProcess(), start, end - start);
67 #elif defined(__arm__) && !defined(__APPLE__)
68 #if defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__)
69   struct arm_sync_icache_args arg;
70 
71   arg.addr = (uintptr_t)start;
72   arg.len = (uintptr_t)end - (uintptr_t)start;
73 
74   sysarch(ARM_SYNC_ICACHE, &arg);
75 #elif defined(__linux__)
76 // We used to include asm/unistd.h for the __ARM_NR_cacheflush define, but
77 // it also brought many other unused defines, as well as a dependency on
78 // kernel headers to be installed.
79 //
80 // This value is stable at least since Linux 3.13 and should remain so for
81 // compatibility reasons, warranting it's re-definition here.
82 #define __ARM_NR_cacheflush 0x0f0002
83   register int start_reg __asm("r0") = (int)(intptr_t)start;
84   const register int end_reg __asm("r1") = (int)(intptr_t)end;
85   const register int flags __asm("r2") = 0;
86   const register int syscall_nr __asm("r7") = __ARM_NR_cacheflush;
87   __asm __volatile("svc 0x0"
88                    : "=r"(start_reg)
89                    : "r"(syscall_nr), "r"(start_reg), "r"(end_reg), "r"(flags));
90   assert(start_reg == 0 && "Cache flush syscall failed.");
91 #else
92   compilerrt_abort();
93 #endif
94 #elif defined(__linux__) && defined(__mips__)
95   const uintptr_t start_int = (uintptr_t)start;
96   const uintptr_t end_int = (uintptr_t)end;
97   syscall(__NR_cacheflush, start, (end_int - start_int), BCACHE);
98 #elif defined(__mips__) && defined(__OpenBSD__)
99   cacheflush(start, (uintptr_t)end - (uintptr_t)start, BCACHE);
100 #elif defined(__aarch64__) && !defined(__APPLE__)
101   uint64_t xstart = (uint64_t)(uintptr_t)start;
102   uint64_t xend = (uint64_t)(uintptr_t)end;
103 
104   // Get Cache Type Info.
105   static uint64_t ctr_el0 = 0;
106   if (ctr_el0 == 0)
107     __asm __volatile("mrs %0, ctr_el0" : "=r"(ctr_el0));
108 
109   // The DC and IC instructions must use 64-bit registers so we don't use
110   // uintptr_t in case this runs in an IPL32 environment.
111   uint64_t addr;
112 
113   // If CTR_EL0.IDC is set, data cache cleaning to the point of unification
114   // is not required for instruction to data coherence.
115   if (((ctr_el0 >> 28) & 0x1) == 0x0) {
116     const size_t dcache_line_size = 4 << ((ctr_el0 >> 16) & 15);
117     for (addr = xstart & ~(dcache_line_size - 1); addr < xend;
118          addr += dcache_line_size)
119       __asm __volatile("dc cvau, %0" ::"r"(addr));
120   }
121   __asm __volatile("dsb ish");
122 
123   // If CTR_EL0.DIC is set, instruction cache invalidation to the point of
124   // unification is not required for instruction to data coherence.
125   if (((ctr_el0 >> 29) & 0x1) == 0x0) {
126     const size_t icache_line_size = 4 << ((ctr_el0 >> 0) & 15);
127     for (addr = xstart & ~(icache_line_size - 1); addr < xend;
128          addr += icache_line_size)
129       __asm __volatile("ic ivau, %0" ::"r"(addr));
130     __asm __volatile("dsb ish");
131   }
132   __asm __volatile("isb sy");
133 #elif defined(__powerpc__)
134   // Newer CPUs have a bigger line size made of multiple blocks, so the
135   // following value is a minimal common denominator for what used to be
136   // a single block cache line and is therefore inneficient.
137   const size_t line_size = 32;
138   const size_t len = (uintptr_t)end - (uintptr_t)start;
139 
140   const uintptr_t mask = ~(line_size - 1);
141   const uintptr_t start_line = ((uintptr_t)start) & mask;
142   const uintptr_t end_line = ((uintptr_t)start + len + line_size - 1) & mask;
143 
144   for (uintptr_t line = start_line; line < end_line; line += line_size)
145     __asm__ volatile("dcbf 0, %0" : : "r"(line));
146   __asm__ volatile("sync");
147 
148   for (uintptr_t line = start_line; line < end_line; line += line_size)
149     __asm__ volatile("icbi 0, %0" : : "r"(line));
150   __asm__ volatile("isync");
151 #elif defined(__sparc__)
152   const size_t dword_size = 8;
153   const size_t len = (uintptr_t)end - (uintptr_t)start;
154 
155   const uintptr_t mask = ~(dword_size - 1);
156   const uintptr_t start_dword = ((uintptr_t)start) & mask;
157   const uintptr_t end_dword = ((uintptr_t)start + len + dword_size - 1) & mask;
158 
159   for (uintptr_t dword = start_dword; dword < end_dword; dword += dword_size)
160     __asm__ volatile("flush %0" : : "r"(dword));
161 #elif defined(__riscv) && defined(__linux__)
162   // See: arch/riscv/include/asm/cacheflush.h, arch/riscv/kernel/sys_riscv.c
163   register void *start_reg __asm("a0") = start;
164   const register void *end_reg __asm("a1") = end;
165   // "0" means that we clear cache for all threads (SYS_RISCV_FLUSH_ICACHE_ALL)
166   const register long flags __asm("a2") = 0;
167   const register long syscall_nr __asm("a7") = __NR_riscv_flush_icache;
168   __asm __volatile("ecall"
169                    : "=r"(start_reg)
170                    : "r"(start_reg), "r"(end_reg), "r"(flags), "r"(syscall_nr));
171   assert(start_reg == 0 && "Cache flush syscall failed.");
172 #elif defined(__riscv) && defined(__OpenBSD__)
173   struct riscv_sync_icache_args arg;
174 
175   arg.addr = (uintptr_t)start;
176   arg.len = (uintptr_t)end - (uintptr_t)start;
177 
178   sysarch(RISCV_SYNC_ICACHE, &arg);
179 #else
180 #if __APPLE__
181   // On Darwin, sys_icache_invalidate() provides this functionality
182   sys_icache_invalidate(start, end - start);
183 #elif defined(__ve__)
184   __asm__ volatile("fencec 2");
185 #else
186   compilerrt_abort();
187 #endif
188 #endif
189 }
190