filename | src/xlat/xltcache.h |
changeset | 1195:072131b61d2a |
prev | 1188:1cc9bb0b3848 |
next | 1214:49152b3d8b75 |
author | nkeynes |
date | Mon Dec 12 21:15:44 2011 +1000 (12 years ago) |
permissions | -rw-r--r-- |
last change | Handle branch delay-slot instruction that falls on the next page correctly. - Generate the right end PC in the first place (sh4trans.c) - Allow blocks to be marked as both entry point + continuation, and specifically handle invalidation when first entry of a page is a continuation == flush previous page as well. |
file | annotate | diff | log | raw |
nkeynes@991 | 1 | /** |
nkeynes@991 | 2 | * $Id$ |
nkeynes@991 | 3 | * |
nkeynes@991 | 4 | * Translation cache support (architecture independent) |
nkeynes@991 | 5 | * |
nkeynes@991 | 6 | * Copyright (c) 2005 Nathan Keynes. |
nkeynes@991 | 7 | * |
nkeynes@991 | 8 | * This program is free software; you can redistribute it and/or modify |
nkeynes@991 | 9 | * it under the terms of the GNU General Public License as published by |
nkeynes@991 | 10 | * the Free Software Foundation; either version 2 of the License, or |
nkeynes@991 | 11 | * (at your option) any later version. |
nkeynes@991 | 12 | * |
nkeynes@991 | 13 | * This program is distributed in the hope that it will be useful, |
nkeynes@991 | 14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
nkeynes@991 | 15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
nkeynes@991 | 16 | * GNU General Public License for more details. |
nkeynes@991 | 17 | */ |
nkeynes@991 | 18 | |
nkeynes@991 | 19 | #include "dream.h" |
nkeynes@991 | 20 | #include "mem.h" |
nkeynes@991 | 21 | |
nkeynes@991 | 22 | #ifndef lxdream_xltcache_H |
nkeynes@991 | 23 | #define lxdream_xltcache_H 1 |
nkeynes@991 | 24 | |
nkeynes@991 | 25 | /** |
nkeynes@991 | 26 | * For now, recovery is purely a matter of mapping native pc => sh4 pc, |
nkeynes@991 | 27 | * and updating sh4r.pc & sh4r.slice_cycles accordingly. In future more |
nkeynes@991 | 28 | * detailed recovery may be required if the translator optimizes more |
nkeynes@991 | 29 | * agressively. |
nkeynes@991 | 30 | * |
nkeynes@991 | 31 | * The recovery table contains (at least) one entry per abortable instruction, |
nkeynes@991 | 32 | * |
nkeynes@991 | 33 | */ |
nkeynes@991 | 34 | typedef struct xlat_recovery_record { |
nkeynes@991 | 35 | uint32_t xlat_offset; // native (translated) pc |
nkeynes@991 | 36 | uint32_t sh4_icount; // instruction number of the corresponding SH4 instruction |
nkeynes@991 | 37 | // (0 = first instruction, 1 = second instruction, ... ) |
nkeynes@991 | 38 | } *xlat_recovery_record_t; |
nkeynes@991 | 39 | |
nkeynes@991 | 40 | struct xlat_cache_block { |
nkeynes@991 | 41 | int active; /* 0 = deleted, 1 = normal. 2 = accessed (temp-space only) */ |
nkeynes@991 | 42 | uint32_t size; |
nkeynes@991 | 43 | void **lut_entry; /* For deletion */ |
nkeynes@1149 | 44 | void *chain; |
nkeynes@1186 | 45 | void *use_list; |
nkeynes@991 | 46 | uint32_t xlat_sh4_mode; /* comparison with sh4r.xlat_sh4_mode */ |
nkeynes@991 | 47 | uint32_t recover_table_offset; // Offset from code[0] of the recovery table; |
nkeynes@991 | 48 | uint32_t recover_table_size; |
nkeynes@991 | 49 | unsigned char code[0]; |
nkeynes@991 | 50 | } __attribute__((packed)); |
nkeynes@991 | 51 | |
nkeynes@991 | 52 | typedef struct xlat_cache_block *xlat_cache_block_t; |
nkeynes@991 | 53 | |
nkeynes@991 | 54 | #define XLAT_BLOCK_FOR_CODE(code) (((xlat_cache_block_t)code)-1) |
nkeynes@991 | 55 | |
nkeynes@1091 | 56 | #define XLAT_BLOCK_MODE(code) (XLAT_BLOCK_FOR_CODE(code)->xlat_sh4_mode) |
nkeynes@1149 | 57 | #define XLAT_BLOCK_CHAIN(code) (XLAT_BLOCK_FOR_CODE(code)->chain) |
nkeynes@1091 | 58 | #define XLAT_RECOVERY_TABLE(code) ((xlat_recovery_record_t)(((char *)code) + XLAT_BLOCK_FOR_CODE(code)->recover_table_offset)) |
nkeynes@991 | 59 | |
nkeynes@991 | 60 | /** |
nkeynes@991 | 61 | * Initialize the translation cache |
nkeynes@991 | 62 | */ |
nkeynes@991 | 63 | void xlat_cache_init(void); |
nkeynes@991 | 64 | |
nkeynes@991 | 65 | /** |
nkeynes@991 | 66 | * Returns the next block in the new cache list that can be written to by the |
nkeynes@991 | 67 | * translator. |
nkeynes@991 | 68 | */ |
nkeynes@991 | 69 | xlat_cache_block_t xlat_start_block(sh4addr_t address); |
nkeynes@991 | 70 | |
nkeynes@991 | 71 | /** |
nkeynes@991 | 72 | * Increases the current block size (only valid between calls to xlat_start_block() |
nkeynes@991 | 73 | * and xlat_commit_block()). |
nkeynes@991 | 74 | * @return the new block, which may be different from the old block. |
nkeynes@991 | 75 | */ |
nkeynes@991 | 76 | xlat_cache_block_t xlat_extend_block( uint32_t newSize ); |
nkeynes@991 | 77 | |
nkeynes@991 | 78 | /** |
nkeynes@991 | 79 | * Commit the current translation block |
nkeynes@991 | 80 | * @param destsize final size of the translation in bytes. |
nkeynes@1195 | 81 | * @param startpc PC at the start of the translation block. |
nkeynes@1195 | 82 | * @param endpc PC at the end of the translation block (i.e. the address of the |
nkeynes@1195 | 83 | * next instruction after the block). |
nkeynes@991 | 84 | */ |
nkeynes@1195 | 85 | void xlat_commit_block( uint32_t destsize, sh4addr_t startpc, sh4addr_t endpc ); |
nkeynes@991 | 86 | |
nkeynes@991 | 87 | /** |
nkeynes@991 | 88 | * Dump the disassembly of the specified code block to a stream |
nkeynes@991 | 89 | * (primarily for debugging purposes) |
nkeynes@991 | 90 | * @param out The stream to write the output to |
nkeynes@991 | 91 | * @param code a translated block |
nkeynes@991 | 92 | */ |
nkeynes@991 | 93 | void xlat_disasm_block( FILE *out, void *code ); |
nkeynes@991 | 94 | |
nkeynes@991 | 95 | |
nkeynes@991 | 96 | /** |
nkeynes@991 | 97 | * Delete (deactivate) the specified block from the cache. Caller is responsible |
nkeynes@991 | 98 | * for ensuring that there really is a block there. |
nkeynes@991 | 99 | */ |
nkeynes@991 | 100 | void xlat_delete_block( xlat_cache_block_t block ); |
nkeynes@991 | 101 | |
nkeynes@991 | 102 | /** |
nkeynes@991 | 103 | * Retrieve the entry point for the translated code corresponding to the given |
nkeynes@991 | 104 | * SH4 address, or NULL if there is no code for that address. |
nkeynes@991 | 105 | */ |
nkeynes@991 | 106 | void * FASTCALL xlat_get_code( sh4addr_t address ); |
nkeynes@991 | 107 | |
nkeynes@991 | 108 | /** |
nkeynes@991 | 109 | * Retrieve the pre-instruction recovery record corresponding to the given |
nkeynes@991 | 110 | * native address, or NULL if there is no recovery code for the address. |
nkeynes@991 | 111 | * @param code The code block containing the recovery table. |
nkeynes@991 | 112 | * @param native_pc A pointer that must be within the currently executing |
nkeynes@991 | 113 | * return the first record before or equal to the given pc. |
nkeynes@991 | 114 | * translation block. |
nkeynes@991 | 115 | */ |
nkeynes@991 | 116 | struct xlat_recovery_record *xlat_get_pre_recovery( void *code, void *native_pc ); |
nkeynes@991 | 117 | |
nkeynes@991 | 118 | |
nkeynes@991 | 119 | /** |
nkeynes@991 | 120 | * Retrieve the entry point for the translated code corresponding to the given |
nkeynes@991 | 121 | * SH4 virtual address, or NULL if there is no code for the address. |
nkeynes@991 | 122 | * If the virtual address cannot be resolved, this method will raise a TLB miss |
nkeynes@991 | 123 | * exception, and return NULL. |
nkeynes@991 | 124 | */ |
nkeynes@991 | 125 | void * FASTCALL xlat_get_code_by_vma( sh4vma_t address ); |
nkeynes@991 | 126 | |
nkeynes@991 | 127 | /** |
nkeynes@991 | 128 | * Retrieve the address of the lookup table entry corresponding to the |
nkeynes@991 | 129 | * given SH4 address. |
nkeynes@991 | 130 | */ |
nkeynes@991 | 131 | void ** FASTCALL xlat_get_lut_entry( sh4addr_t address ); |
nkeynes@991 | 132 | |
nkeynes@991 | 133 | /** |
nkeynes@991 | 134 | * Retrieve the current host address of the running translated code block. |
nkeynes@991 | 135 | * @return the host PC, or null if there is no currently executing translated |
nkeynes@991 | 136 | * block (or the stack is corrupted) |
nkeynes@991 | 137 | * Note: the implementation of this method is host (and calling-convention) specific. |
nkeynes@991 | 138 | * @param block_start start of the block the PC should be in |
nkeynes@991 | 139 | * @param block_size size of the code block in bytes. |
nkeynes@991 | 140 | */ |
nkeynes@991 | 141 | void *xlat_get_native_pc( void *block_start, uint32_t block_size ); |
nkeynes@991 | 142 | |
nkeynes@991 | 143 | /** |
nkeynes@991 | 144 | * Retrieve the size of the block starting at the specified pointer. If the |
nkeynes@991 | 145 | * pointer is not a valid code block, the return value is undefined. |
nkeynes@991 | 146 | */ |
nkeynes@991 | 147 | uint32_t FASTCALL xlat_get_block_size( void *ptr ); |
nkeynes@991 | 148 | |
nkeynes@991 | 149 | /** |
nkeynes@991 | 150 | * Retrieve the size of the code in the block starting at the specified |
nkeynes@991 | 151 | * pointer. Effectively this is xlat_get_block_size() minus the size of |
nkeynes@991 | 152 | * the recovery table. If the pointer is not a valid code block, the |
nkeynes@991 | 153 | * return value is undefined. |
nkeynes@991 | 154 | */ |
nkeynes@991 | 155 | uint32_t FASTCALL xlat_get_code_size( void *ptr ); |
nkeynes@991 | 156 | |
nkeynes@991 | 157 | /** |
nkeynes@991 | 158 | * Flush the code cache for the page containing the given address |
nkeynes@991 | 159 | */ |
nkeynes@991 | 160 | void FASTCALL xlat_flush_page( sh4addr_t address ); |
nkeynes@991 | 161 | |
nkeynes@991 | 162 | void FASTCALL xlat_invalidate_word( sh4addr_t address ); |
nkeynes@991 | 163 | void FASTCALL xlat_invalidate_long( sh4addr_t address ); |
nkeynes@991 | 164 | |
nkeynes@991 | 165 | |
nkeynes@991 | 166 | /** |
nkeynes@991 | 167 | * Invalidate the code cache for a memory region |
nkeynes@991 | 168 | */ |
nkeynes@991 | 169 | void FASTCALL xlat_invalidate_block( sh4addr_t address, size_t bytes ); |
nkeynes@991 | 170 | |
nkeynes@991 | 171 | /** |
nkeynes@991 | 172 | * Flush the entire code cache. This isn't as cheap as one might like |
nkeynes@991 | 173 | */ |
nkeynes@991 | 174 | void xlat_flush_cache(); |
nkeynes@991 | 175 | |
nkeynes@991 | 176 | /** |
nkeynes@1091 | 177 | * Test if the given pointer is within the translation cache, and (is likely) |
nkeynes@1091 | 178 | * the start of a code block |
nkeynes@1091 | 179 | */ |
nkeynes@1091 | 180 | gboolean xlat_is_code_pointer( void *p ); |
nkeynes@1091 | 181 | |
nkeynes@1091 | 182 | /** |
nkeynes@991 | 183 | * Check the internal integrity of the cache |
nkeynes@991 | 184 | */ |
nkeynes@991 | 185 | void xlat_check_integrity(); |
nkeynes@991 | 186 | |
nkeynes@1182 | 187 | /** |
nkeynes@1188 | 188 | * Short record with block + pc, used for activity dumps |
nkeynes@1188 | 189 | */ |
nkeynes@1188 | 190 | typedef struct xlat_block_ref { |
nkeynes@1188 | 191 | xlat_cache_block_t block; |
nkeynes@1188 | 192 | uint32_t pc; |
nkeynes@1188 | 193 | } *xlat_block_ref_t; |
nkeynes@1188 | 194 | |
nkeynes@1188 | 195 | /** |
nkeynes@1188 | 196 | * Fetch the top numRecords translated blocks by number of executions (requires block |
nkeynes@1182 | 197 | * profiling to be turned on in order to give meaningful results). |
nkeynes@1188 | 198 | * @param topN Number of blocks to print. |
nkeynes@1188 | 199 | * @return the number of records retrieved |
nkeynes@1182 | 200 | */ |
nkeynes@1188 | 201 | unsigned int xlat_get_cache_blocks_by_activity( xlat_block_ref_t records, size_t numRecords ); |
nkeynes@1188 | 202 | |
nkeynes@1182 | 203 | void xlat_dump_cache_by_activity( unsigned int topN ); |
nkeynes@1182 | 204 | |
nkeynes@991 | 205 | #endif /* lxdream_xltcache_H */ |
.