cranelift_codegen_meta/shared/
settings.rs

1use crate::cdsl::settings::{SettingGroup, SettingGroupBuilder};
2
3pub(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        "regalloc_algorithm",
35        "Algorithm to use in register allocator.",
36        r#"
37            Supported options:
38
39            - `backtracking`: A backtracking allocator with range splitting; more expensive
40                              but generates better code.
41            - `single_pass`: A single-pass algorithm that yields quick compilation but
42                             results in code with more register spills and moves.
43        "#,
44        vec!["backtracking", "single_pass"],
45    );
46
47    settings.add_enum(
48        "opt_level",
49        "Optimization level for generated code.",
50        r#"
51            Supported levels:
52
53            - `none`: Minimise compile time by disabling most optimizations.
54            - `speed`: Generate the fastest possible code
55            - `speed_and_size`: like "speed", but also perform transformations aimed at reducing code size.
56        "#,
57        vec!["none", "speed", "speed_and_size"],
58    );
59
60    settings.add_bool(
61        "enable_alias_analysis",
62        "Do redundant-load optimizations with alias analysis.",
63        r#"
64            This enables the use of a simple alias analysis to optimize away redundant loads.
65            Only effective when `opt_level` is `speed` or `speed_and_size`.
66        "#,
67        true,
68    );
69
70    settings.add_bool(
71        "enable_verifier",
72        "Run the Cranelift IR verifier at strategic times during compilation.",
73        r#"
74            This makes compilation slower but catches many bugs. The verifier is always enabled by
75            default, which is useful during development.
76        "#,
77        true,
78    );
79
80    settings.add_bool(
81        "enable_pcc",
82        "Enable proof-carrying code translation validation.",
83        r#"
84            This adds a proof-carrying-code mode. Proof-carrying code (PCC) is a strategy to verify
85            that the compiler preserves certain properties or invariants in the compiled code.
86            For example, a frontend that translates WebAssembly to CLIF can embed PCC facts in
87            the CLIF, and Cranelift will verify that the final machine code satisfies the stated
88            facts at each intermediate computed value. Loads and stores can be marked as "checked"
89            and their memory effects can be verified as safe.
90        "#,
91        false,
92    );
93
94    // Note that Cranelift doesn't currently need an is_pie flag, because PIE is
95    // just PIC where symbols can't be pre-empted, which can be expressed with the
96    // `colocated` flag on external functions and global values.
97    settings.add_bool(
98        "is_pic",
99        "Enable Position-Independent Code generation.",
100        "",
101        false,
102    );
103
104    settings.add_bool(
105        "use_colocated_libcalls",
106        "Use colocated libcalls.",
107        r#"
108            Generate code that assumes that libcalls can be declared "colocated",
109            meaning they will be defined along with the current function, such that
110            they can use more efficient addressing.
111        "#,
112        false,
113    );
114
115    settings.add_bool(
116        "enable_nan_canonicalization",
117        "Enable NaN canonicalization.",
118        r#"
119            This replaces NaNs with a single canonical value, for users requiring
120            entirely deterministic WebAssembly computation. This is not required
121            by the WebAssembly spec, so it is not enabled by default.
122        "#,
123        false,
124    );
125
126    settings.add_bool(
127        "enable_pinned_reg",
128        "Enable the use of the pinned register.",
129        r#"
130            This register is excluded from register allocation, and is completely under the control of
131            the end-user. It is possible to read it via the get_pinned_reg instruction, and to set it
132            with the set_pinned_reg instruction.
133        "#,
134        false,
135    );
136
137    settings.add_enum(
138        "tls_model",
139        "Defines the model used to perform TLS accesses.",
140        "",
141        vec!["none", "elf_gd", "macho", "coff"],
142    );
143
144    settings.add_enum(
145        "stack_switch_model",
146        "Defines the model used to performing stack switching.",
147        r#"
148           This determines the compilation of `stack_switch` instructions. If
149           set to `basic`, we simply save all registers, update stack pointer
150           and frame pointer (if needed), and jump to the target IP.
151           If set to `update_windows_tib`, we *additionally* update information
152           about the active stack in Windows' Thread Information Block.
153        "#,
154        vec!["none", "basic", "update_windows_tib"],
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        "enable_multi_ret_implicit_sret",
200        "Enable support for sret arg introduction when there are too many ret vals.",
201        r#"
202            When there are more returns than available return registers, the
203            return value has to be returned through the introduction of a
204            return area pointer. Normally this return area pointer has to be
205            introduced as `ArgumentPurpose::StructReturn` parameter, but for
206            backward compatibility reasons Cranelift also supports implicitly
207            introducing this parameter and writing the return values through it.
208
209            **This option currently does not conform to platform ABIs and the
210            used ABI should not be assumed to remain the same between Cranelift
211            versions.**
212
213            This option is **deprecated** and will be removed in the future.
214
215            Because of the above issues, and complexities of native ABI support
216            for the concept in general, Cranelift's support for multiple return
217            values may also be removed in the future (#9510). For the most
218            robust solution, it is recommended to build a convention on top of
219            Cranelift's primitives for passing multiple return values, for
220            example by allocating a stackslot in the caller, passing it as an
221            explicit StructReturn argument, storing return values in the callee,
222            and loading results in the caller.
223        "#,
224        false,
225    );
226
227    settings.add_bool(
228        "unwind_info",
229        "Generate unwind information.",
230        r#"
231            This increases metadata size and compile time, but allows for the
232            debugger to trace frames, is needed for GC tracing that relies on
233            libunwind (such as in Wasmtime), and is unconditionally needed on
234            certain platforms (such as Windows) that must always be able to unwind.
235          "#,
236        true,
237    );
238
239    settings.add_bool(
240        "preserve_frame_pointers",
241        "Preserve frame pointers",
242        r#"
243            Preserving frame pointers -- even inside leaf functions -- makes it
244            easy to capture the stack of a running program, without requiring any
245            side tables or metadata (like `.eh_frame` sections). Many sampling
246            profilers and similar tools walk frame pointers to capture stacks.
247            Enabling this option will play nice with those tools.
248        "#,
249        false,
250    );
251
252    settings.add_bool(
253        "machine_code_cfg_info",
254        "Generate CFG metadata for machine code.",
255        r#"
256            This increases metadata size and compile time, but allows for the
257            embedder to more easily post-process or analyze the generated
258            machine code. It provides code offsets for the start of each
259            basic block in the generated machine code, and a list of CFG
260            edges (with blocks identified by start offsets) between them.
261            This is useful for, e.g., machine-code analyses that verify certain
262            properties of the generated code.
263        "#,
264        false,
265    );
266
267    // Stack probing options.
268
269    settings.add_bool(
270        "enable_probestack",
271        "Enable the use of stack probes for supported calling conventions.",
272        "",
273        false,
274    );
275
276    settings.add_num(
277        "probestack_size_log2",
278        "The log2 of the size of the stack guard region.",
279        r#"
280            Stack frames larger than this size will have stack overflow checked
281            by calling the probestack function.
282
283            The default is 12, which translates to a size of 4096.
284        "#,
285        12,
286    );
287
288    settings.add_enum(
289        "probestack_strategy",
290        "Controls what kinds of stack probes are emitted.",
291        r#"
292            Supported strategies:
293
294            - `outline`: Always emits stack probes as calls to a probe stack function.
295            - `inline`: Always emits inline stack probes.
296        "#,
297        vec!["outline", "inline"],
298    );
299
300    // Spectre options. (Only read by wasmtime-cranelift)
301    // FIXME move configuration out of Cranelift into Wasmtime
302
303    settings.add_bool(
304        "enable_heap_access_spectre_mitigation",
305        "Enable Spectre mitigation on heap bounds checks.",
306        r#"
307            This is a no-op for any heap that needs no bounds checks; e.g.,
308            if the limit is static and the guard region is large enough that
309            the index cannot reach past it.
310
311            This option is enabled by default because it is highly
312            recommended for secure sandboxing. The embedder should consider
313            the security implications carefully before disabling this option.
314        "#,
315        true,
316    );
317
318    settings.add_bool(
319        "enable_table_access_spectre_mitigation",
320        "Enable Spectre mitigation on table bounds checks.",
321        r#"
322            This option uses a conditional move to ensure that when a table
323            access index is bounds-checked and a conditional branch is used
324            for the out-of-bounds case, a misspeculation of that conditional
325            branch (falsely predicted in-bounds) will select an in-bounds
326            index to load on the speculative path.
327
328            This option is enabled by default because it is highly
329            recommended for secure sandboxing. The embedder should consider
330            the security implications carefully before disabling this option.
331        "#,
332        true,
333    );
334
335    settings.add_bool(
336        "enable_incremental_compilation_cache_checks",
337        "Enable additional checks for debugging the incremental compilation cache.",
338        r#"
339            Enables additional checks that are useful during development of the incremental
340            compilation cache. This should be mostly useful for Cranelift hackers, as well as for
341            helping to debug false incremental cache positives for embedders.
342
343            This option is disabled by default and requires enabling the "incremental-cache" Cargo
344            feature in cranelift-codegen.
345        "#,
346        false,
347    );
348
349    settings.add_num(
350        "bb_padding_log2_minus_one",
351        "The log2 of the size to insert dummy padding between basic blocks",
352        r#"
353            This is a debugging option for stressing various cases during code
354            generation without requiring large functions. This will insert
355            0-byte padding between basic blocks of the specified size.
356
357            The amount of padding inserted two raised to the power of this value
358            minus one. If this value is 0 then no padding is inserted.
359
360            The default for this option is 0 to insert no padding as it's only
361            intended for testing and development.
362        "#,
363        0,
364    );
365
366    settings.add_num(
367        "log2_min_function_alignment",
368        "The log2 of the minimum alignment of functions",
369        "The bigger of this value and the default alignment will be used as actual alignment.",
370        0,
371    );
372
373    // When adding new settings please check if they can also be added
374    // in cranelift/fuzzgen/src/lib.rs for fuzzing.
375    settings.build()
376}