1 use crate::cdsl::settings::{SettingGroup, SettingGroupBuilder}; 2 3 pub(crate) fn define() -> SettingGroup { 4 let mut settings = SettingGroupBuilder::new("shared"); 5 6 settings.add_bool( 7 "regalloc_checker", 8 "Enable the symbolic checker for register allocation.", 9 r#" 10 This performs a verification that the register allocator preserves 11 equivalent dataflow with respect to the original (pre-regalloc) 12 program. This analysis is somewhat expensive. However, if it succeeds, 13 it provides independent evidence (by a carefully-reviewed, from-first-principles 14 analysis) that no regalloc bugs were triggered for the particular compilations 15 performed. This is a valuable assurance to have as regalloc bugs can be 16 very dangerous and difficult to debug. 17 "#, 18 false, 19 ); 20 21 settings.add_bool( 22 "regalloc_verbose_logs", 23 "Enable verbose debug logs for regalloc2.", 24 r#" 25 This adds extra logging for regalloc2 output, that is quite valuable to understand 26 decisions taken by the register allocator as well as debugging it. It is disabled by 27 default, as it can cause many log calls which can slow down compilation by a large 28 amount. 29 "#, 30 false, 31 ); 32 33 settings.add_enum( 34 "opt_level", 35 "Optimization level for generated code.", 36 r#" 37 Supported levels: 38 39 - `none`: Minimise compile time by disabling most optimizations. 40 - `speed`: Generate the fastest possible code 41 - `speed_and_size`: like "speed", but also perform transformations aimed at reducing code size. 42 "#, 43 vec!["none", "speed", "speed_and_size"], 44 ); 45 46 settings.add_bool( 47 "enable_alias_analysis", 48 "Do redundant-load optimizations with alias analysis.", 49 r#" 50 This enables the use of a simple alias analysis to optimize away redundant loads. 51 Only effective when `opt_level` is `speed` or `speed_and_size`. 52 "#, 53 true, 54 ); 55 56 settings.add_bool( 57 "enable_verifier", 58 "Run the Cranelift IR verifier at strategic times during compilation.", 59 r#" 60 This makes compilation slower but catches many bugs. The verifier is always enabled by 61 default, which is useful during development. 62 "#, 63 true, 64 ); 65 66 settings.add_bool( 67 "enable_pcc", 68 "Enable proof-carrying code translation validation.", 69 r#" 70 This adds a proof-carrying-code mode. Proof-carrying code (PCC) is a strategy to verify 71 that the compiler preserves certain properties or invariants in the compiled code. 72 For example, a frontend that translates WebAssembly to CLIF can embed PCC facts in 73 the CLIF, and Cranelift will verify that the final machine code satisfies the stated 74 facts at each intermediate computed value. Loads and stores can be marked as "checked" 75 and their memory effects can be verified as safe. 76 "#, 77 false, 78 ); 79 80 // Note that Cranelift doesn't currently need an is_pie flag, because PIE is 81 // just PIC where symbols can't be pre-empted, which can be expressed with the 82 // `colocated` flag on external functions and global values. 83 settings.add_bool( 84 "is_pic", 85 "Enable Position-Independent Code generation.", 86 "", 87 false, 88 ); 89 90 settings.add_bool( 91 "use_colocated_libcalls", 92 "Use colocated libcalls.", 93 r#" 94 Generate code that assumes that libcalls can be declared "colocated", 95 meaning they will be defined along with the current function, such that 96 they can use more efficient addressing. 97 "#, 98 false, 99 ); 100 101 settings.add_bool( 102 "enable_float", 103 "Enable the use of floating-point instructions.", 104 r#" 105 Disabling use of floating-point instructions is not yet implemented. 106 "#, 107 true, 108 ); 109 110 settings.add_bool( 111 "enable_nan_canonicalization", 112 "Enable NaN canonicalization.", 113 r#" 114 This replaces NaNs with a single canonical value, for users requiring 115 entirely deterministic WebAssembly computation. This is not required 116 by the WebAssembly spec, so it is not enabled by default. 117 "#, 118 false, 119 ); 120 121 settings.add_bool( 122 "enable_pinned_reg", 123 "Enable the use of the pinned register.", 124 r#" 125 This register is excluded from register allocation, and is completely under the control of 126 the end-user. It is possible to read it via the get_pinned_reg instruction, and to set it 127 with the set_pinned_reg instruction. 128 "#, 129 false, 130 ); 131 132 settings.add_bool( 133 "enable_atomics", 134 "Enable the use of atomic instructions", 135 "", 136 true, 137 ); 138 139 settings.add_bool( 140 "enable_safepoints", 141 "Enable safepoint instruction insertions.", 142 r#" 143 This will allow the emit_stack_maps() function to insert the safepoint 144 instruction on top of calls and interrupt traps in order to display the 145 live reference values at that point in the program. 146 "#, 147 false, 148 ); 149 150 settings.add_enum( 151 "tls_model", 152 "Defines the model used to perform TLS accesses.", 153 "", 154 vec!["none", "elf_gd", "macho", "coff"], 155 ); 156 157 settings.add_enum( 158 "libcall_call_conv", 159 "Defines the calling convention to use for LibCalls call expansion.", 160 r#" 161 This may be different from the ISA default calling convention. 162 163 The default value is to use the same calling convention as the ISA 164 default calling convention. 165 166 This list should be kept in sync with the list of calling 167 conventions available in isa/call_conv.rs. 168 "#, 169 vec![ 170 "isa_default", 171 "fast", 172 "cold", 173 "system_v", 174 "windows_fastcall", 175 "apple_aarch64", 176 "probestack", 177 ], 178 ); 179 180 settings.add_bool( 181 "enable_llvm_abi_extensions", 182 "Enable various ABI extensions defined by LLVM's behavior.", 183 r#" 184 In some cases, LLVM's implementation of an ABI (calling convention) 185 goes beyond a standard and supports additional argument types or 186 behavior. This option instructs Cranelift codegen to follow LLVM's 187 behavior where applicable. 188 189 Currently, this applies only to Windows Fastcall on x86-64, and 190 allows an `i128` argument to be spread across two 64-bit integer 191 registers. The Fastcall implementation otherwise does not support 192 `i128` arguments, and will panic if they are present and this 193 option is not set. 194 "#, 195 false, 196 ); 197 198 settings.add_bool( 199 "unwind_info", 200 "Generate unwind information.", 201 r#" 202 This increases metadata size and compile time, but allows for the 203 debugger to trace frames, is needed for GC tracing that relies on 204 libunwind (such as in Wasmtime), and is unconditionally needed on 205 certain platforms (such as Windows) that must always be able to unwind. 206 "#, 207 true, 208 ); 209 210 settings.add_bool( 211 "preserve_frame_pointers", 212 "Preserve frame pointers", 213 r#" 214 Preserving frame pointers -- even inside leaf functions -- makes it 215 easy to capture the stack of a running program, without requiring any 216 side tables or metadata (like `.eh_frame` sections). Many sampling 217 profilers and similar tools walk frame pointers to capture stacks. 218 Enabling this option will play nice with those tools. 219 "#, 220 false, 221 ); 222 223 settings.add_bool( 224 "machine_code_cfg_info", 225 "Generate CFG metadata for machine code.", 226 r#" 227 This increases metadata size and compile time, but allows for the 228 embedder to more easily post-process or analyze the generated 229 machine code. It provides code offsets for the start of each 230 basic block in the generated machine code, and a list of CFG 231 edges (with blocks identified by start offsets) between them. 232 This is useful for, e.g., machine-code analyses that verify certain 233 properties of the generated code. 234 "#, 235 false, 236 ); 237 238 // Stack probing options. 239 240 settings.add_bool( 241 "enable_probestack", 242 "Enable the use of stack probes for supported calling conventions.", 243 "", 244 false, 245 ); 246 247 settings.add_bool( 248 "probestack_func_adjusts_sp", 249 "Enable if the stack probe adjusts the stack pointer.", 250 "", 251 false, 252 ); 253 254 settings.add_num( 255 "probestack_size_log2", 256 "The log2 of the size of the stack guard region.", 257 r#" 258 Stack frames larger than this size will have stack overflow checked 259 by calling the probestack function. 260 261 The default is 12, which translates to a size of 4096. 262 "#, 263 12, 264 ); 265 266 settings.add_enum( 267 "probestack_strategy", 268 "Controls what kinds of stack probes are emitted.", 269 r#" 270 Supported strategies: 271 272 - `outline`: Always emits stack probes as calls to a probe stack function. 273 - `inline`: Always emits inline stack probes. 274 "#, 275 vec!["outline", "inline"], 276 ); 277 278 // Jump table options. 279 280 settings.add_bool( 281 "enable_jump_tables", 282 "Enable the use of jump tables in generated machine code.", 283 "", 284 true, 285 ); 286 287 // Spectre options. 288 289 settings.add_bool( 290 "enable_heap_access_spectre_mitigation", 291 "Enable Spectre mitigation on heap bounds checks.", 292 r#" 293 This is a no-op for any heap that needs no bounds checks; e.g., 294 if the limit is static and the guard region is large enough that 295 the index cannot reach past it. 296 297 This option is enabled by default because it is highly 298 recommended for secure sandboxing. The embedder should consider 299 the security implications carefully before disabling this option. 300 "#, 301 true, 302 ); 303 304 settings.add_bool( 305 "enable_table_access_spectre_mitigation", 306 "Enable Spectre mitigation on table bounds checks.", 307 r#" 308 This option uses a conditional move to ensure that when a table 309 access index is bounds-checked and a conditional branch is used 310 for the out-of-bounds case, a misspeculation of that conditional 311 branch (falsely predicted in-bounds) will select an in-bounds 312 index to load on the speculative path. 313 314 This option is enabled by default because it is highly 315 recommended for secure sandboxing. The embedder should consider 316 the security implications carefully before disabling this option. 317 "#, 318 true, 319 ); 320 321 settings.add_bool( 322 "enable_incremental_compilation_cache_checks", 323 "Enable additional checks for debugging the incremental compilation cache.", 324 r#" 325 Enables additional checks that are useful during development of the incremental 326 compilation cache. This should be mostly useful for Cranelift hackers, as well as for 327 helping to debug false incremental cache positives for embedders. 328 329 This option is disabled by default and requires enabling the "incremental-cache" Cargo 330 feature in cranelift-codegen. 331 "#, 332 false, 333 ); 334 335 settings.add_num( 336 "bb_padding_log2_minus_one", 337 "The log2 of the size to insert dummy padding between basic blocks", 338 r#" 339 This is a debugging option for stressing various cases during code 340 generation without requiring large functions. This will insert 341 0-byte padding between basic blocks of the specified size. 342 343 The amount of padding inserted two raised to the power of this value 344 minus one. If this value is 0 then no padding is inserted. 345 346 The default for this option is 0 to insert no padding as it's only 347 intended for testing and development. 348 "#, 349 0, 350 ); 351 352 // When adding new settings please check if they can also be added 353 // in cranelift/fuzzgen/src/lib.rs for fuzzing. 354 settings.build() 355 } 356