Skip to main content

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    // Note that Cranelift doesn't currently need an is_pie flag, because PIE is
81    // just PIC where symbols can't be preempted, 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_nan_canonicalization",
103        "Enable NaN canonicalization.",
104        r#"
105            This replaces NaNs with a single canonical value, for users requiring
106            entirely deterministic WebAssembly computation. This is not required
107            by the WebAssembly spec, so it is not enabled by default.
108        "#,
109        false,
110    );
111
112    settings.add_bool(
113        "enable_pinned_reg",
114        "Enable the use of the pinned register.",
115        r#"
116            This register is excluded from register allocation, and is completely under the control of
117            the end-user. It is possible to read it via the get_pinned_reg instruction, and to set it
118            with the set_pinned_reg instruction.
119        "#,
120        false,
121    );
122
123    settings.add_enum(
124        "tls_model",
125        "Defines the model used to perform TLS accesses.",
126        "",
127        vec!["none", "elf_gd", "macho", "coff"],
128    );
129
130    settings.add_enum(
131        "stack_switch_model",
132        "Defines the model used to performing stack switching.",
133        r#"
134           This determines the compilation of `stack_switch` instructions. If
135           set to `basic`, we simply save all registers, update stack pointer
136           and frame pointer (if needed), and jump to the target IP.
137           If set to `update_windows_tib`, we *additionally* update information
138           about the active stack in Windows' Thread Information Block.
139        "#,
140        vec!["none", "basic", "update_windows_tib"],
141    );
142
143    settings.add_enum(
144        "libcall_call_conv",
145        "Defines the calling convention to use for LibCalls call expansion.",
146        r#"
147            This may be different from the ISA default calling convention.
148
149            The default value is to use the same calling convention as the ISA
150            default calling convention.
151
152            This list should be kept in sync with the list of calling
153            conventions available in isa/call_conv.rs.
154        "#,
155        vec![
156            "isa_default",
157            "fast",
158            "system_v",
159            "windows_fastcall",
160            "apple_aarch64",
161            "probestack",
162            "preserve_all",
163        ],
164    );
165
166    settings.add_bool(
167        "enable_llvm_abi_extensions",
168        "Enable various ABI extensions defined by LLVM's behavior.",
169        r#"
170            In some cases, LLVM's implementation of an ABI (calling convention)
171            goes beyond a standard and supports additional argument types or
172            behavior. This option instructs Cranelift codegen to follow LLVM's
173            behavior where applicable.
174
175            Currently, this applies only to Windows Fastcall on x86-64, and
176            allows an `i128` argument to be spread across two 64-bit integer
177            registers. The Fastcall implementation otherwise does not support
178            `i128` arguments, and will panic if they are present and this
179            option is not set.
180        "#,
181        false,
182    );
183
184    settings.add_bool(
185        "enable_multi_ret_implicit_sret",
186        "Enable support for sret arg introduction when there are too many ret vals.",
187        r#"
188            When there are more returns than available return registers, the
189            return value has to be returned through the introduction of a
190            return area pointer. Normally this return area pointer has to be
191            introduced as `ArgumentPurpose::StructReturn` parameter, but for
192            backward compatibility reasons Cranelift also supports implicitly
193            introducing this parameter and writing the return values through it.
194
195            **This option currently does not conform to platform ABIs and the
196            used ABI should not be assumed to remain the same between Cranelift
197            versions.**
198
199            This option is **deprecated** and will be removed in the future.
200
201            Because of the above issues, and complexities of native ABI support
202            for the concept in general, Cranelift's support for multiple return
203            values may also be removed in the future (#9510). For the most
204            robust solution, it is recommended to build a convention on top of
205            Cranelift's primitives for passing multiple return values, for
206            example by allocating a stackslot in the caller, passing it as an
207            explicit StructReturn argument, storing return values in the callee,
208            and loading results in the caller.
209        "#,
210        false,
211    );
212
213    settings.add_bool(
214        "unwind_info",
215        "Generate unwind information.",
216        r#"
217            This increases metadata size and compile time, but allows for the
218            debugger to trace frames, is needed for GC tracing that relies on
219            libunwind (such as in Wasmtime), and is unconditionally needed on
220            certain platforms (such as Windows) that must always be able to unwind.
221          "#,
222        true,
223    );
224
225    settings.add_bool(
226        "preserve_frame_pointers",
227        "Preserve frame pointers",
228        r#"
229            Preserving frame pointers -- even inside leaf functions -- makes it
230            easy to capture the stack of a running program, without requiring any
231            side tables or metadata (like `.eh_frame` sections). Many sampling
232            profilers and similar tools walk frame pointers to capture stacks.
233            Enabling this option will play nice with those tools.
234        "#,
235        false,
236    );
237
238    settings.add_bool(
239        "machine_code_cfg_info",
240        "Generate CFG metadata for machine code.",
241        r#"
242            This increases metadata size and compile time, but allows for the
243            embedder to more easily post-process or analyze the generated
244            machine code. It provides code offsets for the start of each
245            basic block in the generated machine code, and a list of CFG
246            edges (with blocks identified by start offsets) between them.
247            This is useful for, e.g., machine-code analyses that verify certain
248            properties of the generated code.
249        "#,
250        false,
251    );
252
253    // Stack probing options.
254
255    settings.add_bool(
256        "enable_probestack",
257        "Enable the use of stack probes for supported calling conventions.",
258        "",
259        false,
260    );
261
262    settings.add_num(
263        "probestack_size_log2",
264        "The log2 of the size of the stack guard region.",
265        r#"
266            Stack frames larger than this size will have stack overflow checked
267            by calling the probestack function.
268
269            The default is 12, which translates to a size of 4096.
270        "#,
271        12,
272    );
273
274    settings.add_enum(
275        "probestack_strategy",
276        "Controls what kinds of stack probes are emitted.",
277        r#"
278            Supported strategies:
279
280            - `outline`: Always emits stack probes as calls to a probe stack function.
281            - `inline`: Always emits inline stack probes.
282        "#,
283        vec!["outline", "inline"],
284    );
285
286    // Spectre options. (Only read by wasmtime-cranelift)
287    // FIXME move configuration out of Cranelift into Wasmtime
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    settings.add_num(
353        "log2_min_function_alignment",
354        "The log2 of the minimum alignment of functions",
355        "The bigger of this value and the default alignment will be used as actual alignment.",
356        0,
357    );
358
359    // When adding new settings please check if they can also be added
360    // in cranelift/fuzzgen/src/lib.rs for fuzzing.
361    settings.build()
362}