xref: /freebsd/contrib/llvm-project/llvm/include/llvm/Target/TargetOptions.h (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 //===-- llvm/Target/TargetOptions.h - Target Options ------------*- C++ -*-===//
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 // This file defines command line option flags that are shared across various
10 // targets.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #ifndef LLVM_TARGET_TARGETOPTIONS_H
15 #define LLVM_TARGET_TARGETOPTIONS_H
16 
17 #include "llvm/ADT/FloatingPointMode.h"
18 #include "llvm/MC/MCTargetOptions.h"
19 
20 #include <memory>
21 
22 namespace llvm {
23   struct fltSemantics;
24   class MachineFunction;
25   class MemoryBuffer;
26 
27   namespace FloatABI {
28     enum ABIType {
29       Default, // Target-specific (either soft or hard depending on triple, etc).
30       Soft,    // Soft float.
31       Hard     // Hard float.
32     };
33   }
34 
35   namespace FPOpFusion {
36     enum FPOpFusionMode {
37       Fast,     // Enable fusion of FP ops wherever it's profitable.
38       Standard, // Only allow fusion of 'blessed' ops (currently just fmuladd).
39       Strict    // Never fuse FP-ops.
40     };
41   }
42 
43   namespace JumpTable {
44     enum JumpTableType {
45       Single,          // Use a single table for all indirect jumptable calls.
46       Arity,           // Use one table per number of function parameters.
47       Simplified,      // Use one table per function type, with types projected
48                        // into 4 types: pointer to non-function, struct,
49                        // primitive, and function pointer.
50       Full             // Use one table per unique function type
51     };
52   }
53 
54   namespace ThreadModel {
55     enum Model {
56       POSIX,  // POSIX Threads
57       Single  // Single Threaded Environment
58     };
59   }
60 
61   enum class BasicBlockSection {
62     All,    // Use Basic Block Sections for all basic blocks.  A section
63             // for every basic block can significantly bloat object file sizes.
64     List,   // Get list of functions & BBs from a file. Selectively enables
65             // basic block sections for a subset of basic blocks which can be
66             // used to control object size bloats from creating sections.
67     Labels, // Do not use Basic Block Sections but label basic blocks.  This
68             // is useful when associating profile counts from virtual addresses
69             // to basic blocks.
70     Preset, // Similar to list but the blocks are identified by passes which
71             // seek to use Basic Block Sections, e.g. MachineFunctionSplitter.
72             // This option cannot be set via the command line.
73     None    // Do not use Basic Block Sections.
74   };
75 
76   enum class EABI {
77     Unknown,
78     Default, // Default means not specified
79     EABI4,   // Target-specific (either 4, 5 or gnu depending on triple).
80     EABI5,
81     GNU
82   };
83 
84   /// Identify a debugger for "tuning" the debug info.
85   ///
86   /// The "debugger tuning" concept allows us to present a more intuitive
87   /// interface that unpacks into different sets of defaults for the various
88   /// individual feature-flag settings, that suit the preferences of the
89   /// various debuggers.  However, it's worth remembering that debuggers are
90   /// not the only consumers of debug info, and some variations in DWARF might
91   /// better be treated as target/platform issues. Fundamentally,
92   /// o if the feature is useful (or not) to a particular debugger, regardless
93   ///   of the target, that's a tuning decision;
94   /// o if the feature is useful (or not) on a particular platform, regardless
95   ///   of the debugger, that's a target decision.
96   /// It's not impossible to see both factors in some specific case.
97   enum class DebuggerKind {
98     Default, ///< No specific tuning requested.
99     GDB,     ///< Tune debug info for gdb.
100     LLDB,    ///< Tune debug info for lldb.
101     SCE,     ///< Tune debug info for SCE targets (e.g. PS4).
102     DBX      ///< Tune debug info for dbx.
103   };
104 
105   /// Enable abort calls when global instruction selection fails to lower/select
106   /// an instruction.
107   enum class GlobalISelAbortMode {
108     Disable,        // Disable the abort.
109     Enable,         // Enable the abort.
110     DisableWithDiag // Disable the abort but emit a diagnostic on failure.
111   };
112 
113   /// Indicates when and how the Swift async frame pointer bit should be set.
114   enum class SwiftAsyncFramePointerMode {
115     /// Determine whether to set the bit statically or dynamically based
116     /// on the deployment target.
117     DeploymentBased,
118     /// Always set the bit.
119     Always,
120     /// Never set the bit.
121     Never,
122   };
123 
124   /// \brief Enumeration value for AMDGPU code object version, which is the
125   /// code object version times 100.
126   enum CodeObjectVersionKind {
127     COV_None,
128     COV_2 = 200, // Unsupported.
129     COV_3 = 300, // Unsupported.
130     COV_4 = 400,
131     COV_5 = 500,
132     COV_6 = 600,
133   };
134 
135   class TargetOptions {
136   public:
TargetOptions()137     TargetOptions()
138         : UnsafeFPMath(false), NoInfsFPMath(false), NoNaNsFPMath(false),
139           NoTrappingFPMath(true), NoSignedZerosFPMath(false),
140           ApproxFuncFPMath(false), EnableAIXExtendedAltivecABI(false),
141           HonorSignDependentRoundingFPMathOption(false), NoZerosInBSS(false),
142           GuaranteedTailCallOpt(false), StackSymbolOrdering(true),
143           EnableFastISel(false), EnableGlobalISel(false), UseInitArray(false),
144           DisableIntegratedAS(false), FunctionSections(false),
145           DataSections(false), IgnoreXCOFFVisibility(false),
146           XCOFFTracebackTable(true), UniqueSectionNames(true),
147           UniqueBasicBlockSectionNames(false), SeparateNamedSections(false),
148           TrapUnreachable(false), NoTrapAfterNoreturn(false), TLSSize(0),
149           EmulatedTLS(false), EnableTLSDESC(false), EnableIPRA(false),
150           EmitStackSizeSection(false), EnableMachineOutliner(false),
151           EnableMachineFunctionSplitter(false), SupportsDefaultOutlining(false),
152           EmitAddrsig(false), BBAddrMap(false), EmitCallSiteInfo(false),
153           SupportsDebugEntryValues(false), EnableDebugEntryValues(false),
154           ValueTrackingVariableLocations(false), ForceDwarfFrameSection(false),
155           XRayFunctionIndex(true), DebugStrictDwarf(false), Hotpatch(false),
156           PPCGenScalarMASSEntries(false), JMCInstrument(false),
157           EnableCFIFixup(false), MisExpect(false), XCOFFReadOnlyPointers(false),
158           FPDenormalMode(DenormalMode::IEEE, DenormalMode::IEEE) {}
159 
160     /// DisableFramePointerElim - This returns true if frame pointer elimination
161     /// optimization should be disabled for the given machine function.
162     bool DisableFramePointerElim(const MachineFunction &MF) const;
163 
164     /// FramePointerIsReserved - This returns true if the frame pointer must
165     /// always either point to a new frame record or be un-modified in the given
166     /// function.
167     bool FramePointerIsReserved(const MachineFunction &MF) const;
168 
169     /// If greater than 0, override the default value of
170     /// MCAsmInfo::BinutilsVersion.
171     std::pair<int, int> BinutilsVersion{0, 0};
172 
173     /// UnsafeFPMath - This flag is enabled when the
174     /// -enable-unsafe-fp-math flag is specified on the command line.  When
175     /// this flag is off (the default), the code generator is not allowed to
176     /// produce results that are "less precise" than IEEE allows.  This includes
177     /// use of X86 instructions like FSIN and FCOS instead of libcalls.
178     unsigned UnsafeFPMath : 1;
179 
180     /// NoInfsFPMath - This flag is enabled when the
181     /// -enable-no-infs-fp-math flag is specified on the command line. When
182     /// this flag is off (the default), the code generator is not allowed to
183     /// assume the FP arithmetic arguments and results are never +-Infs.
184     unsigned NoInfsFPMath : 1;
185 
186     /// NoNaNsFPMath - This flag is enabled when the
187     /// -enable-no-nans-fp-math flag is specified on the command line. When
188     /// this flag is off (the default), the code generator is not allowed to
189     /// assume the FP arithmetic arguments and results are never NaNs.
190     unsigned NoNaNsFPMath : 1;
191 
192     /// NoTrappingFPMath - This flag is enabled when the
193     /// -enable-no-trapping-fp-math is specified on the command line. This
194     /// specifies that there are no trap handlers to handle exceptions.
195     unsigned NoTrappingFPMath : 1;
196 
197     /// NoSignedZerosFPMath - This flag is enabled when the
198     /// -enable-no-signed-zeros-fp-math is specified on the command line. This
199     /// specifies that optimizations are allowed to treat the sign of a zero
200     /// argument or result as insignificant.
201     unsigned NoSignedZerosFPMath : 1;
202 
203     /// ApproxFuncFPMath - This flag is enabled when the
204     /// -enable-approx-func-fp-math is specified on the command line. This
205     /// specifies that optimizations are allowed to substitute math functions
206     /// with approximate calculations
207     unsigned ApproxFuncFPMath : 1;
208 
209     /// EnableAIXExtendedAltivecABI - This flag returns true when -vec-extabi is
210     /// specified. The code generator is then able to use both volatile and
211     /// nonvolitle vector registers. When false, the code generator only uses
212     /// volatile vector registers which is the default setting on AIX.
213     unsigned EnableAIXExtendedAltivecABI : 1;
214 
215     /// HonorSignDependentRoundingFPMath - This returns true when the
216     /// -enable-sign-dependent-rounding-fp-math is specified.  If this returns
217     /// false (the default), the code generator is allowed to assume that the
218     /// rounding behavior is the default (round-to-zero for all floating point
219     /// to integer conversions, and round-to-nearest for all other arithmetic
220     /// truncations).  If this is enabled (set to true), the code generator must
221     /// assume that the rounding mode may dynamically change.
222     unsigned HonorSignDependentRoundingFPMathOption : 1;
223     bool HonorSignDependentRoundingFPMath() const;
224 
225     /// NoZerosInBSS - By default some codegens place zero-initialized data to
226     /// .bss section. This flag disables such behaviour (necessary, e.g. for
227     /// crt*.o compiling).
228     unsigned NoZerosInBSS : 1;
229 
230     /// GuaranteedTailCallOpt - This flag is enabled when -tailcallopt is
231     /// specified on the commandline. When the flag is on, participating targets
232     /// will perform tail call optimization on all calls which use the fastcc
233     /// calling convention and which satisfy certain target-independent
234     /// criteria (being at the end of a function, having the same return type
235     /// as their parent function, etc.), using an alternate ABI if necessary.
236     unsigned GuaranteedTailCallOpt : 1;
237 
238     /// StackSymbolOrdering - When true, this will allow CodeGen to order
239     /// the local stack symbols (for code size, code locality, or any other
240     /// heuristics). When false, the local symbols are left in whatever order
241     /// they were generated. Default is true.
242     unsigned StackSymbolOrdering : 1;
243 
244     /// EnableFastISel - This flag enables fast-path instruction selection
245     /// which trades away generated code quality in favor of reducing
246     /// compile time.
247     unsigned EnableFastISel : 1;
248 
249     /// EnableGlobalISel - This flag enables global instruction selection.
250     unsigned EnableGlobalISel : 1;
251 
252     /// EnableGlobalISelAbort - Control abort behaviour when global instruction
253     /// selection fails to lower/select an instruction.
254     GlobalISelAbortMode GlobalISelAbort = GlobalISelAbortMode::Enable;
255 
256     /// Control when and how the Swift async frame pointer bit should
257     /// be set.
258     SwiftAsyncFramePointerMode SwiftAsyncFramePointer =
259         SwiftAsyncFramePointerMode::Always;
260 
261     /// UseInitArray - Use .init_array instead of .ctors for static
262     /// constructors.
263     unsigned UseInitArray : 1;
264 
265     /// Disable the integrated assembler.
266     unsigned DisableIntegratedAS : 1;
267 
268     /// Emit functions into separate sections.
269     unsigned FunctionSections : 1;
270 
271     /// Emit data into separate sections.
272     unsigned DataSections : 1;
273 
274     /// Do not emit visibility attribute for xcoff.
275     unsigned IgnoreXCOFFVisibility : 1;
276 
277     /// Emit XCOFF traceback table.
278     unsigned XCOFFTracebackTable : 1;
279 
280     unsigned UniqueSectionNames : 1;
281 
282     /// Use unique names for basic block sections.
283     unsigned UniqueBasicBlockSectionNames : 1;
284 
285     /// Emit named sections with the same name into different sections.
286     unsigned SeparateNamedSections : 1;
287 
288     /// Emit target-specific trap instruction for 'unreachable' IR instructions.
289     unsigned TrapUnreachable : 1;
290 
291     /// Do not emit a trap instruction for 'unreachable' IR instructions behind
292     /// noreturn calls, even if TrapUnreachable is true.
293     unsigned NoTrapAfterNoreturn : 1;
294 
295     /// Bit size of immediate TLS offsets (0 == use the default).
296     unsigned TLSSize : 8;
297 
298     /// EmulatedTLS - This flag enables emulated TLS model, using emutls
299     /// function in the runtime library..
300     unsigned EmulatedTLS : 1;
301 
302     /// EnableTLSDESC - This flag enables TLS Descriptors.
303     unsigned EnableTLSDESC : 1;
304 
305     /// This flag enables InterProcedural Register Allocation (IPRA).
306     unsigned EnableIPRA : 1;
307 
308     /// Emit section containing metadata on function stack sizes.
309     unsigned EmitStackSizeSection : 1;
310 
311     /// Enables the MachineOutliner pass.
312     unsigned EnableMachineOutliner : 1;
313 
314     /// Enables the MachineFunctionSplitter pass.
315     unsigned EnableMachineFunctionSplitter : 1;
316 
317     /// Set if the target supports default outlining behaviour.
318     unsigned SupportsDefaultOutlining : 1;
319 
320     /// Emit address-significance table.
321     unsigned EmitAddrsig : 1;
322 
323     // Emit the SHT_LLVM_BB_ADDR_MAP section containing basic block address
324     // which can be used to map virtual addresses to machine basic blocks.
325     unsigned BBAddrMap : 1;
326 
327     /// Emit basic blocks into separate sections.
328     BasicBlockSection BBSections = BasicBlockSection::None;
329 
330     /// Memory Buffer that contains information on sampled basic blocks and used
331     /// to selectively generate basic block sections.
332     std::shared_ptr<MemoryBuffer> BBSectionsFuncListBuf;
333 
334     /// The flag enables call site info production. It is used only for debug
335     /// info, and it is restricted only to optimized code. This can be used for
336     /// something else, so that should be controlled in the frontend.
337     unsigned EmitCallSiteInfo : 1;
338     /// Set if the target supports the debug entry values by default.
339     unsigned SupportsDebugEntryValues : 1;
340     /// When set to true, the EnableDebugEntryValues option forces production
341     /// of debug entry values even if the target does not officially support
342     /// it. Useful for testing purposes only. This flag should never be checked
343     /// directly, always use \ref ShouldEmitDebugEntryValues instead.
344      unsigned EnableDebugEntryValues : 1;
345     /// NOTE: There are targets that still do not support the debug entry values
346     /// production.
347     bool ShouldEmitDebugEntryValues() const;
348 
349     // When set to true, use experimental new debug variable location tracking,
350     // which seeks to follow the values of variables rather than their location,
351     // post isel.
352     unsigned ValueTrackingVariableLocations : 1;
353 
354     /// Emit DWARF debug frame section.
355     unsigned ForceDwarfFrameSection : 1;
356 
357     /// Emit XRay Function Index section
358     unsigned XRayFunctionIndex : 1;
359 
360     /// When set to true, don't use DWARF extensions in later DWARF versions.
361     /// By default, it is set to false.
362     unsigned DebugStrictDwarf : 1;
363 
364     /// Emit the hotpatch flag in CodeView debug.
365     unsigned Hotpatch : 1;
366 
367     /// Enables scalar MASS conversions
368     unsigned PPCGenScalarMASSEntries : 1;
369 
370     /// Enable JustMyCode instrumentation.
371     unsigned JMCInstrument : 1;
372 
373     /// Enable the CFIFixup pass.
374     unsigned EnableCFIFixup : 1;
375 
376     /// When set to true, enable MisExpect Diagnostics
377     /// By default, it is set to false
378     unsigned MisExpect : 1;
379 
380     /// When set to true, const objects with relocatable address values are put
381     /// into the RO data section.
382     unsigned XCOFFReadOnlyPointers : 1;
383 
384     /// Name of the stack usage file (i.e., .su file) if user passes
385     /// -fstack-usage. If empty, it can be implied that -fstack-usage is not
386     /// passed on the command line.
387     std::string StackUsageOutput;
388 
389     /// If greater than 0, override TargetLoweringBase::PrefLoopAlignment.
390     unsigned LoopAlignment = 0;
391 
392     /// FloatABIType - This setting is set by -float-abi=xxx option is specfied
393     /// on the command line. This setting may either be Default, Soft, or Hard.
394     /// Default selects the target's default behavior. Soft selects the ABI for
395     /// software floating point, but does not indicate that FP hardware may not
396     /// be used. Such a combination is unfortunately popular (e.g.
397     /// arm-apple-darwin). Hard presumes that the normal FP ABI is used.
398     FloatABI::ABIType FloatABIType = FloatABI::Default;
399 
400     /// AllowFPOpFusion - This flag is set by the -fp-contract=xxx option.
401     /// This controls the creation of fused FP ops that store intermediate
402     /// results in higher precision than IEEE allows (E.g. FMAs).
403     ///
404     /// Fast mode - allows formation of fused FP ops whenever they're
405     /// profitable.
406     /// Standard mode - allow fusion only for 'blessed' FP ops. At present the
407     /// only blessed op is the fmuladd intrinsic. In the future more blessed ops
408     /// may be added.
409     /// Strict mode - allow fusion only if/when it can be proven that the excess
410     /// precision won't effect the result.
411     ///
412     /// Note: This option only controls formation of fused ops by the
413     /// optimizers.  Fused operations that are explicitly specified (e.g. FMA
414     /// via the llvm.fma.* intrinsic) will always be honored, regardless of
415     /// the value of this option.
416     FPOpFusion::FPOpFusionMode AllowFPOpFusion = FPOpFusion::Standard;
417 
418     /// ThreadModel - This flag specifies the type of threading model to assume
419     /// for things like atomics
420     ThreadModel::Model ThreadModel = ThreadModel::POSIX;
421 
422     /// EABIVersion - This flag specifies the EABI version
423     EABI EABIVersion = EABI::Default;
424 
425     /// Which debugger to tune for.
426     DebuggerKind DebuggerTuning = DebuggerKind::Default;
427 
428   private:
429     /// Flushing mode to assume in default FP environment.
430     DenormalMode FPDenormalMode;
431 
432     /// Flushing mode to assume in default FP environment, for float/vector of
433     /// float.
434     DenormalMode FP32DenormalMode;
435 
436   public:
setFPDenormalMode(DenormalMode Mode)437     void setFPDenormalMode(DenormalMode Mode) {
438       FPDenormalMode = Mode;
439     }
440 
setFP32DenormalMode(DenormalMode Mode)441     void setFP32DenormalMode(DenormalMode Mode) {
442       FP32DenormalMode = Mode;
443     }
444 
getRawFPDenormalMode()445     DenormalMode getRawFPDenormalMode() const {
446       return FPDenormalMode;
447     }
448 
getRawFP32DenormalMode()449     DenormalMode getRawFP32DenormalMode() const {
450       return FP32DenormalMode;
451     }
452 
453     DenormalMode getDenormalMode(const fltSemantics &FPType) const;
454 
455     /// What exception model to use
456     ExceptionHandling ExceptionModel = ExceptionHandling::None;
457 
458     /// Machine level options.
459     MCTargetOptions MCOptions;
460 
461     /// Stores the filename/path of the final .o/.obj file, to be written in the
462     /// debug information. This is used for emitting the CodeView S_OBJNAME
463     /// record.
464     std::string ObjectFilenameForDebug;
465   };
466 
467 } // End llvm namespace
468 
469 #endif
470