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(__loongarch__) 95 __asm__ volatile("ibar 0"); 96 #elif defined(__mips__) 97 const uintptr_t start_int = (uintptr_t)start; 98 const uintptr_t end_int = (uintptr_t)end; 99 uintptr_t synci_step; 100 __asm__ volatile("rdhwr %0, $1" : "=r"(synci_step)); 101 if (synci_step != 0) { 102 #if __mips_isa_rev >= 6 103 for (uintptr_t p = start_int; p < end_int; p += synci_step) 104 __asm__ volatile("synci 0(%0)" : : "r"(p)); 105 106 // The last "move $at, $0" is the target of jr.hb instead of delay slot. 107 __asm__ volatile(".set noat\n" 108 "sync\n" 109 "addiupc $at, 12\n" 110 "jr.hb $at\n" 111 "move $at, $0\n" 112 ".set at"); 113 #elif defined(__linux__) || defined(__OpenBSD__) 114 // Pre-R6 may not be globalized. And some implementations may give strange 115 // synci_step. So, let's use libc call for it. 116 _flush_cache(start, end_int - start_int, BCACHE); 117 #else 118 (void)start_int; 119 (void)end_int; 120 compilerrt_abort(); 121 #endif 122 } 123 #elif defined(__aarch64__) && !defined(__APPLE__) 124 uint64_t xstart = (uint64_t)(uintptr_t)start; 125 uint64_t xend = (uint64_t)(uintptr_t)end; 126 127 // Get Cache Type Info. 128 static uint64_t ctr_el0 = 0; 129 if (ctr_el0 == 0) 130 __asm __volatile("mrs %0, ctr_el0" : "=r"(ctr_el0)); 131 132 // The DC and IC instructions must use 64-bit registers so we don't use 133 // uintptr_t in case this runs in an IPL32 environment. 134 uint64_t addr; 135 136 // If CTR_EL0.IDC is set, data cache cleaning to the point of unification 137 // is not required for instruction to data coherence. 138 if (((ctr_el0 >> 28) & 0x1) == 0x0) { 139 const size_t dcache_line_size = 4 << ((ctr_el0 >> 16) & 15); 140 for (addr = xstart & ~(dcache_line_size - 1); addr < xend; 141 addr += dcache_line_size) 142 __asm __volatile("dc cvau, %0" ::"r"(addr)); 143 } 144 __asm __volatile("dsb ish"); 145 146 // If CTR_EL0.DIC is set, instruction cache invalidation to the point of 147 // unification is not required for instruction to data coherence. 148 if (((ctr_el0 >> 29) & 0x1) == 0x0) { 149 const size_t icache_line_size = 4 << ((ctr_el0 >> 0) & 15); 150 for (addr = xstart & ~(icache_line_size - 1); addr < xend; 151 addr += icache_line_size) 152 __asm __volatile("ic ivau, %0" ::"r"(addr)); 153 __asm __volatile("dsb ish"); 154 } 155 __asm __volatile("isb sy"); 156 #elif defined(__powerpc__) 157 // Newer CPUs have a bigger line size made of multiple blocks, so the 158 // following value is a minimal common denominator for what used to be 159 // a single block cache line and is therefore inneficient. 160 const size_t line_size = 32; 161 const size_t len = (uintptr_t)end - (uintptr_t)start; 162 163 const uintptr_t mask = ~(line_size - 1); 164 const uintptr_t start_line = ((uintptr_t)start) & mask; 165 const uintptr_t end_line = ((uintptr_t)start + len + line_size - 1) & mask; 166 167 for (uintptr_t line = start_line; line < end_line; line += line_size) 168 __asm__ volatile("dcbf 0, %0" : : "r"(line)); 169 __asm__ volatile("sync"); 170 171 for (uintptr_t line = start_line; line < end_line; line += line_size) 172 __asm__ volatile("icbi 0, %0" : : "r"(line)); 173 __asm__ volatile("isync"); 174 #elif defined(__sparc__) 175 const size_t dword_size = 8; 176 const size_t len = (uintptr_t)end - (uintptr_t)start; 177 178 const uintptr_t mask = ~(dword_size - 1); 179 const uintptr_t start_dword = ((uintptr_t)start) & mask; 180 const uintptr_t end_dword = ((uintptr_t)start + len + dword_size - 1) & mask; 181 182 for (uintptr_t dword = start_dword; dword < end_dword; dword += dword_size) 183 __asm__ volatile("flush %0" : : "r"(dword)); 184 #elif defined(__riscv) && defined(__linux__) 185 // See: arch/riscv/include/asm/cacheflush.h, arch/riscv/kernel/sys_riscv.c 186 register void *start_reg __asm("a0") = start; 187 const register void *end_reg __asm("a1") = end; 188 // "0" means that we clear cache for all threads (SYS_RISCV_FLUSH_ICACHE_ALL) 189 const register long flags __asm("a2") = 0; 190 const register long syscall_nr __asm("a7") = __NR_riscv_flush_icache; 191 __asm __volatile("ecall" 192 : "=r"(start_reg) 193 : "r"(start_reg), "r"(end_reg), "r"(flags), "r"(syscall_nr)); 194 assert(start_reg == 0 && "Cache flush syscall failed."); 195 #elif defined(__riscv) && defined(__OpenBSD__) 196 struct riscv_sync_icache_args arg; 197 198 arg.addr = (uintptr_t)start; 199 arg.len = (uintptr_t)end - (uintptr_t)start; 200 201 sysarch(RISCV_SYNC_ICACHE, &arg); 202 #elif defined(__ve__) 203 __asm__ volatile("fencec 2"); 204 #else 205 #if __APPLE__ 206 // On Darwin, sys_icache_invalidate() provides this functionality 207 sys_icache_invalidate(start, end - start); 208 #else 209 compilerrt_abort(); 210 #endif 211 #endif 212 } 213