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   };
133 
134   class TargetOptions {
135   public:
TargetOptions()136     TargetOptions()
137         : UnsafeFPMath(false), NoInfsFPMath(false), NoNaNsFPMath(false),
138           NoTrappingFPMath(true), NoSignedZerosFPMath(false),
139           ApproxFuncFPMath(false), EnableAIXExtendedAltivecABI(false),
140           HonorSignDependentRoundingFPMathOption(false), NoZerosInBSS(false),
141           GuaranteedTailCallOpt(false), StackSymbolOrdering(true),
142           EnableFastISel(false), EnableGlobalISel(false), UseInitArray(false),
143           DisableIntegratedAS(false), RelaxELFRelocations(true),
144           FunctionSections(false), DataSections(false),
145           IgnoreXCOFFVisibility(false), XCOFFTracebackTable(true),
146           UniqueSectionNames(true), UniqueBasicBlockSectionNames(false),
147           TrapUnreachable(false), NoTrapAfterNoreturn(false), TLSSize(0),
148           EmulatedTLS(false), EnableTLSDESC(false), EnableIPRA(false),
149           EmitStackSizeSection(false), EnableMachineOutliner(false),
150           EnableMachineFunctionSplitter(false), SupportsDefaultOutlining(false),
151           EmitAddrsig(false), EmitCallSiteInfo(false),
152           SupportsDebugEntryValues(false), EnableDebugEntryValues(false),
153           ValueTrackingVariableLocations(false), ForceDwarfFrameSection(false),
154           XRayFunctionIndex(true), DebugStrictDwarf(false), Hotpatch(false),
155           PPCGenScalarMASSEntries(false), JMCInstrument(false),
156           EnableCFIFixup(false), MisExpect(false), XCOFFReadOnlyPointers(false),
157           FPDenormalMode(DenormalMode::IEEE, DenormalMode::IEEE) {}
158 
159     /// DisableFramePointerElim - This returns true if frame pointer elimination
160     /// optimization should be disabled for the given machine function.
161     bool DisableFramePointerElim(const MachineFunction &MF) const;
162 
163     /// If greater than 0, override the default value of
164     /// MCAsmInfo::BinutilsVersion.
165     std::pair<int, int> BinutilsVersion{0, 0};
166 
167     /// UnsafeFPMath - This flag is enabled when the
168     /// -enable-unsafe-fp-math flag is specified on the command line.  When
169     /// this flag is off (the default), the code generator is not allowed to
170     /// produce results that are "less precise" than IEEE allows.  This includes
171     /// use of X86 instructions like FSIN and FCOS instead of libcalls.
172     unsigned UnsafeFPMath : 1;
173 
174     /// NoInfsFPMath - This flag is enabled when the
175     /// -enable-no-infs-fp-math flag is specified on the command line. When
176     /// this flag is off (the default), the code generator is not allowed to
177     /// assume the FP arithmetic arguments and results are never +-Infs.
178     unsigned NoInfsFPMath : 1;
179 
180     /// NoNaNsFPMath - This flag is enabled when the
181     /// -enable-no-nans-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 NaNs.
184     unsigned NoNaNsFPMath : 1;
185 
186     /// NoTrappingFPMath - This flag is enabled when the
187     /// -enable-no-trapping-fp-math is specified on the command line. This
188     /// specifies that there are no trap handlers to handle exceptions.
189     unsigned NoTrappingFPMath : 1;
190 
191     /// NoSignedZerosFPMath - This flag is enabled when the
192     /// -enable-no-signed-zeros-fp-math is specified on the command line. This
193     /// specifies that optimizations are allowed to treat the sign of a zero
194     /// argument or result as insignificant.
195     unsigned NoSignedZerosFPMath : 1;
196 
197     /// ApproxFuncFPMath - This flag is enabled when the
198     /// -enable-approx-func-fp-math is specified on the command line. This
199     /// specifies that optimizations are allowed to substitute math functions
200     /// with approximate calculations
201     unsigned ApproxFuncFPMath : 1;
202 
203     /// EnableAIXExtendedAltivecABI - This flag returns true when -vec-extabi is
204     /// specified. The code generator is then able to use both volatile and
205     /// nonvolitle vector registers. When false, the code generator only uses
206     /// volatile vector registers which is the default setting on AIX.
207     unsigned EnableAIXExtendedAltivecABI : 1;
208 
209     /// HonorSignDependentRoundingFPMath - This returns true when the
210     /// -enable-sign-dependent-rounding-fp-math is specified.  If this returns
211     /// false (the default), the code generator is allowed to assume that the
212     /// rounding behavior is the default (round-to-zero for all floating point
213     /// to integer conversions, and round-to-nearest for all other arithmetic
214     /// truncations).  If this is enabled (set to true), the code generator must
215     /// assume that the rounding mode may dynamically change.
216     unsigned HonorSignDependentRoundingFPMathOption : 1;
217     bool HonorSignDependentRoundingFPMath() const;
218 
219     /// NoZerosInBSS - By default some codegens place zero-initialized data to
220     /// .bss section. This flag disables such behaviour (necessary, e.g. for
221     /// crt*.o compiling).
222     unsigned NoZerosInBSS : 1;
223 
224     /// GuaranteedTailCallOpt - This flag is enabled when -tailcallopt is
225     /// specified on the commandline. When the flag is on, participating targets
226     /// will perform tail call optimization on all calls which use the fastcc
227     /// calling convention and which satisfy certain target-independent
228     /// criteria (being at the end of a function, having the same return type
229     /// as their parent function, etc.), using an alternate ABI if necessary.
230     unsigned GuaranteedTailCallOpt : 1;
231 
232     /// StackSymbolOrdering - When true, this will allow CodeGen to order
233     /// the local stack symbols (for code size, code locality, or any other
234     /// heuristics). When false, the local symbols are left in whatever order
235     /// they were generated. Default is true.
236     unsigned StackSymbolOrdering : 1;
237 
238     /// EnableFastISel - This flag enables fast-path instruction selection
239     /// which trades away generated code quality in favor of reducing
240     /// compile time.
241     unsigned EnableFastISel : 1;
242 
243     /// EnableGlobalISel - This flag enables global instruction selection.
244     unsigned EnableGlobalISel : 1;
245 
246     /// EnableGlobalISelAbort - Control abort behaviour when global instruction
247     /// selection fails to lower/select an instruction.
248     GlobalISelAbortMode GlobalISelAbort = GlobalISelAbortMode::Enable;
249 
250     /// Control when and how the Swift async frame pointer bit should
251     /// be set.
252     SwiftAsyncFramePointerMode SwiftAsyncFramePointer =
253         SwiftAsyncFramePointerMode::Always;
254 
255     /// UseInitArray - Use .init_array instead of .ctors for static
256     /// constructors.
257     unsigned UseInitArray : 1;
258 
259     /// Disable the integrated assembler.
260     unsigned DisableIntegratedAS : 1;
261 
262     /// Compress DWARF debug sections.
263     DebugCompressionType CompressDebugSections = DebugCompressionType::None;
264 
265     unsigned RelaxELFRelocations : 1;
266 
267     /// Emit functions into separate sections.
268     unsigned FunctionSections : 1;
269 
270     /// Emit data into separate sections.
271     unsigned DataSections : 1;
272 
273     /// Do not emit visibility attribute for xcoff.
274     unsigned IgnoreXCOFFVisibility : 1;
275 
276     /// Emit XCOFF traceback table.
277     unsigned XCOFFTracebackTable : 1;
278 
279     unsigned UniqueSectionNames : 1;
280 
281     /// Use unique names for basic block sections.
282     unsigned UniqueBasicBlockSectionNames : 1;
283 
284     /// Emit target-specific trap instruction for 'unreachable' IR instructions.
285     unsigned TrapUnreachable : 1;
286 
287     /// Do not emit a trap instruction for 'unreachable' IR instructions behind
288     /// noreturn calls, even if TrapUnreachable is true.
289     unsigned NoTrapAfterNoreturn : 1;
290 
291     /// Bit size of immediate TLS offsets (0 == use the default).
292     unsigned TLSSize : 8;
293 
294     /// EmulatedTLS - This flag enables emulated TLS model, using emutls
295     /// function in the runtime library..
296     unsigned EmulatedTLS : 1;
297 
298     /// EnableTLSDESC - This flag enables TLS Descriptors.
299     unsigned EnableTLSDESC : 1;
300 
301     /// This flag enables InterProcedural Register Allocation (IPRA).
302     unsigned EnableIPRA : 1;
303 
304     /// Emit section containing metadata on function stack sizes.
305     unsigned EmitStackSizeSection : 1;
306 
307     /// Enables the MachineOutliner pass.
308     unsigned EnableMachineOutliner : 1;
309 
310     /// Enables the MachineFunctionSplitter pass.
311     unsigned EnableMachineFunctionSplitter : 1;
312 
313     /// Set if the target supports default outlining behaviour.
314     unsigned SupportsDefaultOutlining : 1;
315 
316     /// Emit address-significance table.
317     unsigned EmitAddrsig : 1;
318 
319     /// Emit basic blocks into separate sections.
320     BasicBlockSection BBSections = BasicBlockSection::None;
321 
322     /// Memory Buffer that contains information on sampled basic blocks and used
323     /// to selectively generate basic block sections.
324     std::shared_ptr<MemoryBuffer> BBSectionsFuncListBuf;
325 
326     /// The flag enables call site info production. It is used only for debug
327     /// info, and it is restricted only to optimized code. This can be used for
328     /// something else, so that should be controlled in the frontend.
329     unsigned EmitCallSiteInfo : 1;
330     /// Set if the target supports the debug entry values by default.
331     unsigned SupportsDebugEntryValues : 1;
332     /// When set to true, the EnableDebugEntryValues option forces production
333     /// of debug entry values even if the target does not officially support
334     /// it. Useful for testing purposes only. This flag should never be checked
335     /// directly, always use \ref ShouldEmitDebugEntryValues instead.
336      unsigned EnableDebugEntryValues : 1;
337     /// NOTE: There are targets that still do not support the debug entry values
338     /// production.
339     bool ShouldEmitDebugEntryValues() const;
340 
341     // When set to true, use experimental new debug variable location tracking,
342     // which seeks to follow the values of variables rather than their location,
343     // post isel.
344     unsigned ValueTrackingVariableLocations : 1;
345 
346     /// Emit DWARF debug frame section.
347     unsigned ForceDwarfFrameSection : 1;
348 
349     /// Emit XRay Function Index section
350     unsigned XRayFunctionIndex : 1;
351 
352     /// When set to true, don't use DWARF extensions in later DWARF versions.
353     /// By default, it is set to false.
354     unsigned DebugStrictDwarf : 1;
355 
356     /// Emit the hotpatch flag in CodeView debug.
357     unsigned Hotpatch : 1;
358 
359     /// Enables scalar MASS conversions
360     unsigned PPCGenScalarMASSEntries : 1;
361 
362     /// Enable JustMyCode instrumentation.
363     unsigned JMCInstrument : 1;
364 
365     /// Enable the CFIFixup pass.
366     unsigned EnableCFIFixup : 1;
367 
368     /// When set to true, enable MisExpect Diagnostics
369     /// By default, it is set to false
370     unsigned MisExpect : 1;
371 
372     /// When set to true, const objects with relocatable address values are put
373     /// into the RO data section.
374     unsigned XCOFFReadOnlyPointers : 1;
375 
376     /// Name of the stack usage file (i.e., .su file) if user passes
377     /// -fstack-usage. If empty, it can be implied that -fstack-usage is not
378     /// passed on the command line.
379     std::string StackUsageOutput;
380 
381     /// If greater than 0, override TargetLoweringBase::PrefLoopAlignment.
382     unsigned LoopAlignment = 0;
383 
384     /// FloatABIType - This setting is set by -float-abi=xxx option is specfied
385     /// on the command line. This setting may either be Default, Soft, or Hard.
386     /// Default selects the target's default behavior. Soft selects the ABI for
387     /// software floating point, but does not indicate that FP hardware may not
388     /// be used. Such a combination is unfortunately popular (e.g.
389     /// arm-apple-darwin). Hard presumes that the normal FP ABI is used.
390     FloatABI::ABIType FloatABIType = FloatABI::Default;
391 
392     /// AllowFPOpFusion - This flag is set by the -fp-contract=xxx option.
393     /// This controls the creation of fused FP ops that store intermediate
394     /// results in higher precision than IEEE allows (E.g. FMAs).
395     ///
396     /// Fast mode - allows formation of fused FP ops whenever they're
397     /// profitable.
398     /// Standard mode - allow fusion only for 'blessed' FP ops. At present the
399     /// only blessed op is the fmuladd intrinsic. In the future more blessed ops
400     /// may be added.
401     /// Strict mode - allow fusion only if/when it can be proven that the excess
402     /// precision won't effect the result.
403     ///
404     /// Note: This option only controls formation of fused ops by the
405     /// optimizers.  Fused operations that are explicitly specified (e.g. FMA
406     /// via the llvm.fma.* intrinsic) will always be honored, regardless of
407     /// the value of this option.
408     FPOpFusion::FPOpFusionMode AllowFPOpFusion = FPOpFusion::Standard;
409 
410     /// ThreadModel - This flag specifies the type of threading model to assume
411     /// for things like atomics
412     ThreadModel::Model ThreadModel = ThreadModel::POSIX;
413 
414     /// EABIVersion - This flag specifies the EABI version
415     EABI EABIVersion = EABI::Default;
416 
417     /// Which debugger to tune for.
418     DebuggerKind DebuggerTuning = DebuggerKind::Default;
419 
420   private:
421     /// Flushing mode to assume in default FP environment.
422     DenormalMode FPDenormalMode;
423 
424     /// Flushing mode to assume in default FP environment, for float/vector of
425     /// float.
426     DenormalMode FP32DenormalMode;
427 
428   public:
setFPDenormalMode(DenormalMode Mode)429     void setFPDenormalMode(DenormalMode Mode) {
430       FPDenormalMode = Mode;
431     }
432 
setFP32DenormalMode(DenormalMode Mode)433     void setFP32DenormalMode(DenormalMode Mode) {
434       FP32DenormalMode = Mode;
435     }
436 
getRawFPDenormalMode()437     DenormalMode getRawFPDenormalMode() const {
438       return FPDenormalMode;
439     }
440 
getRawFP32DenormalMode()441     DenormalMode getRawFP32DenormalMode() const {
442       return FP32DenormalMode;
443     }
444 
445     DenormalMode getDenormalMode(const fltSemantics &FPType) const;
446 
447     /// What exception model to use
448     ExceptionHandling ExceptionModel = ExceptionHandling::None;
449 
450     /// Machine level options.
451     MCTargetOptions MCOptions;
452 
453     /// Stores the filename/path of the final .o/.obj file, to be written in the
454     /// debug information. This is used for emitting the CodeView S_OBJNAME
455     /// record.
456     std::string ObjectFilenameForDebug;
457   };
458 
459 } // End llvm namespace
460 
461 #endif
462