[ARM] MVE compare vector splat combine
[llvm-complete.git] / lib / Target / PowerPC / PPCTargetMachine.cpp
blob213e2f557278661124272ba936c1d04e287c3340
1 //===-- PPCTargetMachine.cpp - Define TargetMachine for PowerPC -----------===//
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 // Top-level implementation for the PowerPC target.
11 //===----------------------------------------------------------------------===//
13 #include "PPCTargetMachine.h"
14 #include "MCTargetDesc/PPCMCTargetDesc.h"
15 #include "PPC.h"
16 #include "PPCMachineScheduler.h"
17 #include "PPCSubtarget.h"
18 #include "PPCTargetObjectFile.h"
19 #include "PPCTargetTransformInfo.h"
20 #include "TargetInfo/PowerPCTargetInfo.h"
21 #include "llvm/ADT/Optional.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/ADT/StringRef.h"
24 #include "llvm/ADT/Triple.h"
25 #include "llvm/Analysis/TargetTransformInfo.h"
26 #include "llvm/CodeGen/Passes.h"
27 #include "llvm/CodeGen/TargetPassConfig.h"
28 #include "llvm/CodeGen/MachineScheduler.h"
29 #include "llvm/IR/Attributes.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/Function.h"
32 #include "llvm/Pass.h"
33 #include "llvm/Support/CodeGen.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/TargetRegistry.h"
36 #include "llvm/Target/TargetLoweringObjectFile.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/Transforms/Scalar.h"
39 #include <cassert>
40 #include <memory>
41 #include <string>
43 using namespace llvm;
46 static cl::opt<bool>
47 EnableBranchCoalescing("enable-ppc-branch-coalesce", cl::Hidden,
48 cl::desc("enable coalescing of duplicate branches for PPC"));
49 static cl::
50 opt<bool> DisableCTRLoops("disable-ppc-ctrloops", cl::Hidden,
51 cl::desc("Disable CTR loops for PPC"));
53 static cl::
54 opt<bool> DisablePreIncPrep("disable-ppc-preinc-prep", cl::Hidden,
55 cl::desc("Disable PPC loop preinc prep"));
57 static cl::opt<bool>
58 VSXFMAMutateEarly("schedule-ppc-vsx-fma-mutation-early",
59 cl::Hidden, cl::desc("Schedule VSX FMA instruction mutation early"));
61 static cl::
62 opt<bool> DisableVSXSwapRemoval("disable-ppc-vsx-swap-removal", cl::Hidden,
63 cl::desc("Disable VSX Swap Removal for PPC"));
65 static cl::
66 opt<bool> DisableQPXLoadSplat("disable-ppc-qpx-load-splat", cl::Hidden,
67 cl::desc("Disable QPX load splat simplification"));
69 static cl::
70 opt<bool> DisableMIPeephole("disable-ppc-peephole", cl::Hidden,
71 cl::desc("Disable machine peepholes for PPC"));
73 static cl::opt<bool>
74 EnableGEPOpt("ppc-gep-opt", cl::Hidden,
75 cl::desc("Enable optimizations on complex GEPs"),
76 cl::init(true));
78 static cl::opt<bool>
79 EnablePrefetch("enable-ppc-prefetching",
80 cl::desc("disable software prefetching on PPC"),
81 cl::init(false), cl::Hidden);
83 static cl::opt<bool>
84 EnableExtraTOCRegDeps("enable-ppc-extra-toc-reg-deps",
85 cl::desc("Add extra TOC register dependencies"),
86 cl::init(true), cl::Hidden);
88 static cl::opt<bool>
89 EnableMachineCombinerPass("ppc-machine-combiner",
90 cl::desc("Enable the machine combiner pass"),
91 cl::init(true), cl::Hidden);
93 static cl::opt<bool>
94 ReduceCRLogical("ppc-reduce-cr-logicals",
95 cl::desc("Expand eligible cr-logical binary ops to branches"),
96 cl::init(false), cl::Hidden);
97 extern "C" void LLVMInitializePowerPCTarget() {
98 // Register the targets
99 RegisterTargetMachine<PPCTargetMachine> A(getThePPC32Target());
100 RegisterTargetMachine<PPCTargetMachine> B(getThePPC64Target());
101 RegisterTargetMachine<PPCTargetMachine> C(getThePPC64LETarget());
103 PassRegistry &PR = *PassRegistry::getPassRegistry();
104 #ifndef NDEBUG
105 initializePPCCTRLoopsVerifyPass(PR);
106 #endif
107 initializePPCLoopPreIncPrepPass(PR);
108 initializePPCTOCRegDepsPass(PR);
109 initializePPCEarlyReturnPass(PR);
110 initializePPCVSXCopyPass(PR);
111 initializePPCVSXFMAMutatePass(PR);
112 initializePPCVSXSwapRemovalPass(PR);
113 initializePPCReduceCRLogicalsPass(PR);
114 initializePPCBSelPass(PR);
115 initializePPCBranchCoalescingPass(PR);
116 initializePPCQPXLoadSplatPass(PR);
117 initializePPCBoolRetToIntPass(PR);
118 initializePPCExpandISELPass(PR);
119 initializePPCPreEmitPeepholePass(PR);
120 initializePPCTLSDynamicCallPass(PR);
121 initializePPCMIPeepholePass(PR);
124 /// Return the datalayout string of a subtarget.
125 static std::string getDataLayoutString(const Triple &T) {
126 bool is64Bit = T.getArch() == Triple::ppc64 || T.getArch() == Triple::ppc64le;
127 std::string Ret;
129 // Most PPC* platforms are big endian, PPC64LE is little endian.
130 if (T.getArch() == Triple::ppc64le)
131 Ret = "e";
132 else
133 Ret = "E";
135 Ret += DataLayout::getManglingComponent(T);
137 // PPC32 has 32 bit pointers. The PS3 (OS Lv2) is a PPC64 machine with 32 bit
138 // pointers.
139 if (!is64Bit || T.getOS() == Triple::Lv2)
140 Ret += "-p:32:32";
142 // Note, the alignment values for f64 and i64 on ppc64 in Darwin
143 // documentation are wrong; these are correct (i.e. "what gcc does").
144 if (is64Bit || !T.isOSDarwin())
145 Ret += "-i64:64";
146 else
147 Ret += "-f64:32:64";
149 // PPC64 has 32 and 64 bit registers, PPC32 has only 32 bit ones.
150 if (is64Bit)
151 Ret += "-n32:64";
152 else
153 Ret += "-n32";
155 return Ret;
158 static std::string computeFSAdditions(StringRef FS, CodeGenOpt::Level OL,
159 const Triple &TT) {
160 std::string FullFS = FS;
162 // Make sure 64-bit features are available when CPUname is generic
163 if (TT.getArch() == Triple::ppc64 || TT.getArch() == Triple::ppc64le) {
164 if (!FullFS.empty())
165 FullFS = "+64bit," + FullFS;
166 else
167 FullFS = "+64bit";
170 if (OL >= CodeGenOpt::Default) {
171 if (!FullFS.empty())
172 FullFS = "+crbits," + FullFS;
173 else
174 FullFS = "+crbits";
177 if (OL != CodeGenOpt::None) {
178 if (!FullFS.empty())
179 FullFS = "+invariant-function-descriptors," + FullFS;
180 else
181 FullFS = "+invariant-function-descriptors";
184 return FullFS;
187 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
188 if (TT.isOSDarwin())
189 return llvm::make_unique<TargetLoweringObjectFileMachO>();
191 if (TT.isOSAIX())
192 return llvm::make_unique<TargetLoweringObjectFileXCOFF>();
194 return llvm::make_unique<PPC64LinuxTargetObjectFile>();
197 static PPCTargetMachine::PPCABI computeTargetABI(const Triple &TT,
198 const TargetOptions &Options) {
199 if (TT.isOSDarwin())
200 report_fatal_error("Darwin is no longer supported for PowerPC");
202 if (Options.MCOptions.getABIName().startswith("elfv1"))
203 return PPCTargetMachine::PPC_ABI_ELFv1;
204 else if (Options.MCOptions.getABIName().startswith("elfv2"))
205 return PPCTargetMachine::PPC_ABI_ELFv2;
207 assert(Options.MCOptions.getABIName().empty() &&
208 "Unknown target-abi option!");
210 if (TT.isMacOSX())
211 return PPCTargetMachine::PPC_ABI_UNKNOWN;
213 switch (TT.getArch()) {
214 case Triple::ppc64le:
215 return PPCTargetMachine::PPC_ABI_ELFv2;
216 case Triple::ppc64:
217 if (TT.getEnvironment() == llvm::Triple::ELFv2)
218 return PPCTargetMachine::PPC_ABI_ELFv2;
219 return PPCTargetMachine::PPC_ABI_ELFv1;
220 default:
221 return PPCTargetMachine::PPC_ABI_UNKNOWN;
225 static Reloc::Model getEffectiveRelocModel(const Triple &TT,
226 Optional<Reloc::Model> RM) {
227 if (RM.hasValue())
228 return *RM;
230 // Darwin defaults to dynamic-no-pic.
231 if (TT.isOSDarwin())
232 return Reloc::DynamicNoPIC;
234 // Big Endian PPC is PIC by default.
235 if (TT.getArch() == Triple::ppc64)
236 return Reloc::PIC_;
238 // Rest are static by default.
239 return Reloc::Static;
242 static CodeModel::Model getEffectivePPCCodeModel(const Triple &TT,
243 Optional<CodeModel::Model> CM,
244 bool JIT) {
245 if (CM) {
246 if (*CM == CodeModel::Tiny)
247 report_fatal_error("Target does not support the tiny CodeModel", false);
248 if (*CM == CodeModel::Kernel)
249 report_fatal_error("Target does not support the kernel CodeModel", false);
250 return *CM;
252 if (!TT.isOSDarwin() && !JIT &&
253 (TT.getArch() == Triple::ppc64 || TT.getArch() == Triple::ppc64le))
254 return CodeModel::Medium;
255 return CodeModel::Small;
259 static ScheduleDAGInstrs *createPPCMachineScheduler(MachineSchedContext *C) {
260 const PPCSubtarget &ST = C->MF->getSubtarget<PPCSubtarget>();
261 ScheduleDAGMILive *DAG =
262 new ScheduleDAGMILive(C, ST.usePPCPreRASchedStrategy() ?
263 llvm::make_unique<PPCPreRASchedStrategy>(C) :
264 llvm::make_unique<GenericScheduler>(C));
265 // add DAG Mutations here.
266 DAG->addMutation(createCopyConstrainDAGMutation(DAG->TII, DAG->TRI));
267 return DAG;
270 static ScheduleDAGInstrs *createPPCPostMachineScheduler(
271 MachineSchedContext *C) {
272 const PPCSubtarget &ST = C->MF->getSubtarget<PPCSubtarget>();
273 ScheduleDAGMI *DAG =
274 new ScheduleDAGMI(C, ST.usePPCPostRASchedStrategy() ?
275 llvm::make_unique<PPCPostRASchedStrategy>(C) :
276 llvm::make_unique<PostGenericScheduler>(C), true);
277 // add DAG Mutations here.
278 return DAG;
281 // The FeatureString here is a little subtle. We are modifying the feature
282 // string with what are (currently) non-function specific overrides as it goes
283 // into the LLVMTargetMachine constructor and then using the stored value in the
284 // Subtarget constructor below it.
285 PPCTargetMachine::PPCTargetMachine(const Target &T, const Triple &TT,
286 StringRef CPU, StringRef FS,
287 const TargetOptions &Options,
288 Optional<Reloc::Model> RM,
289 Optional<CodeModel::Model> CM,
290 CodeGenOpt::Level OL, bool JIT)
291 : LLVMTargetMachine(T, getDataLayoutString(TT), TT, CPU,
292 computeFSAdditions(FS, OL, TT), Options,
293 getEffectiveRelocModel(TT, RM),
294 getEffectivePPCCodeModel(TT, CM, JIT), OL),
295 TLOF(createTLOF(getTargetTriple())),
296 TargetABI(computeTargetABI(TT, Options)) {
297 initAsmInfo();
300 PPCTargetMachine::~PPCTargetMachine() = default;
302 const PPCSubtarget *
303 PPCTargetMachine::getSubtargetImpl(const Function &F) const {
304 Attribute CPUAttr = F.getFnAttribute("target-cpu");
305 Attribute FSAttr = F.getFnAttribute("target-features");
307 std::string CPU = !CPUAttr.hasAttribute(Attribute::None)
308 ? CPUAttr.getValueAsString().str()
309 : TargetCPU;
310 std::string FS = !FSAttr.hasAttribute(Attribute::None)
311 ? FSAttr.getValueAsString().str()
312 : TargetFS;
314 // FIXME: This is related to the code below to reset the target options,
315 // we need to know whether or not the soft float flag is set on the
316 // function before we can generate a subtarget. We also need to use
317 // it as a key for the subtarget since that can be the only difference
318 // between two functions.
319 bool SoftFloat =
320 F.getFnAttribute("use-soft-float").getValueAsString() == "true";
321 // If the soft float attribute is set on the function turn on the soft float
322 // subtarget feature.
323 if (SoftFloat)
324 FS += FS.empty() ? "-hard-float" : ",-hard-float";
326 auto &I = SubtargetMap[CPU + FS];
327 if (!I) {
328 // This needs to be done before we create a new subtarget since any
329 // creation will depend on the TM and the code generation flags on the
330 // function that reside in TargetOptions.
331 resetTargetOptions(F);
332 I = llvm::make_unique<PPCSubtarget>(
333 TargetTriple, CPU,
334 // FIXME: It would be good to have the subtarget additions here
335 // not necessary. Anything that turns them on/off (overrides) ends
336 // up being put at the end of the feature string, but the defaults
337 // shouldn't require adding them. Fixing this means pulling Feature64Bit
338 // out of most of the target cpus in the .td file and making it set only
339 // as part of initialization via the TargetTriple.
340 computeFSAdditions(FS, getOptLevel(), getTargetTriple()), *this);
342 return I.get();
345 //===----------------------------------------------------------------------===//
346 // Pass Pipeline Configuration
347 //===----------------------------------------------------------------------===//
349 namespace {
351 /// PPC Code Generator Pass Configuration Options.
352 class PPCPassConfig : public TargetPassConfig {
353 public:
354 PPCPassConfig(PPCTargetMachine &TM, PassManagerBase &PM)
355 : TargetPassConfig(TM, PM) {
356 // At any optimization level above -O0 we use the Machine Scheduler and not
357 // the default Post RA List Scheduler.
358 if (TM.getOptLevel() != CodeGenOpt::None)
359 substitutePass(&PostRASchedulerID, &PostMachineSchedulerID);
362 PPCTargetMachine &getPPCTargetMachine() const {
363 return getTM<PPCTargetMachine>();
366 void addIRPasses() override;
367 bool addPreISel() override;
368 bool addILPOpts() override;
369 bool addInstSelector() override;
370 void addMachineSSAOptimization() override;
371 void addPreRegAlloc() override;
372 void addPreSched2() override;
373 void addPreEmitPass() override;
374 ScheduleDAGInstrs *
375 createMachineScheduler(MachineSchedContext *C) const override {
376 return createPPCMachineScheduler(C);
378 ScheduleDAGInstrs *
379 createPostMachineScheduler(MachineSchedContext *C) const override {
380 return createPPCPostMachineScheduler(C);
384 } // end anonymous namespace
386 TargetPassConfig *PPCTargetMachine::createPassConfig(PassManagerBase &PM) {
387 return new PPCPassConfig(*this, PM);
390 void PPCPassConfig::addIRPasses() {
391 if (TM->getOptLevel() != CodeGenOpt::None)
392 addPass(createPPCBoolRetToIntPass());
393 addPass(createAtomicExpandPass());
395 // For the BG/Q (or if explicitly requested), add explicit data prefetch
396 // intrinsics.
397 bool UsePrefetching = TM->getTargetTriple().getVendor() == Triple::BGQ &&
398 getOptLevel() != CodeGenOpt::None;
399 if (EnablePrefetch.getNumOccurrences() > 0)
400 UsePrefetching = EnablePrefetch;
401 if (UsePrefetching)
402 addPass(createLoopDataPrefetchPass());
404 if (TM->getOptLevel() >= CodeGenOpt::Default && EnableGEPOpt) {
405 // Call SeparateConstOffsetFromGEP pass to extract constants within indices
406 // and lower a GEP with multiple indices to either arithmetic operations or
407 // multiple GEPs with single index.
408 addPass(createSeparateConstOffsetFromGEPPass(true));
409 // Call EarlyCSE pass to find and remove subexpressions in the lowered
410 // result.
411 addPass(createEarlyCSEPass());
412 // Do loop invariant code motion in case part of the lowered result is
413 // invariant.
414 addPass(createLICMPass());
417 TargetPassConfig::addIRPasses();
420 bool PPCPassConfig::addPreISel() {
421 if (!DisablePreIncPrep && getOptLevel() != CodeGenOpt::None)
422 addPass(createPPCLoopPreIncPrepPass(getPPCTargetMachine()));
424 if (!DisableCTRLoops && getOptLevel() != CodeGenOpt::None)
425 addPass(createHardwareLoopsPass());
427 return false;
430 bool PPCPassConfig::addILPOpts() {
431 addPass(&EarlyIfConverterID);
433 if (EnableMachineCombinerPass)
434 addPass(&MachineCombinerID);
436 return true;
439 bool PPCPassConfig::addInstSelector() {
440 // Install an instruction selector.
441 addPass(createPPCISelDag(getPPCTargetMachine(), getOptLevel()));
443 #ifndef NDEBUG
444 if (!DisableCTRLoops && getOptLevel() != CodeGenOpt::None)
445 addPass(createPPCCTRLoopsVerify());
446 #endif
448 addPass(createPPCVSXCopyPass());
449 return false;
452 void PPCPassConfig::addMachineSSAOptimization() {
453 // PPCBranchCoalescingPass need to be done before machine sinking
454 // since it merges empty blocks.
455 if (EnableBranchCoalescing && getOptLevel() != CodeGenOpt::None)
456 addPass(createPPCBranchCoalescingPass());
457 TargetPassConfig::addMachineSSAOptimization();
458 // For little endian, remove where possible the vector swap instructions
459 // introduced at code generation to normalize vector element order.
460 if (TM->getTargetTriple().getArch() == Triple::ppc64le &&
461 !DisableVSXSwapRemoval)
462 addPass(createPPCVSXSwapRemovalPass());
463 // Reduce the number of cr-logical ops.
464 if (ReduceCRLogical && getOptLevel() != CodeGenOpt::None)
465 addPass(createPPCReduceCRLogicalsPass());
466 // Target-specific peephole cleanups performed after instruction
467 // selection.
468 if (!DisableMIPeephole) {
469 addPass(createPPCMIPeepholePass());
470 addPass(&DeadMachineInstructionElimID);
474 void PPCPassConfig::addPreRegAlloc() {
475 if (getOptLevel() != CodeGenOpt::None) {
476 initializePPCVSXFMAMutatePass(*PassRegistry::getPassRegistry());
477 insertPass(VSXFMAMutateEarly ? &RegisterCoalescerID : &MachineSchedulerID,
478 &PPCVSXFMAMutateID);
481 // FIXME: We probably don't need to run these for -fPIE.
482 if (getPPCTargetMachine().isPositionIndependent()) {
483 // FIXME: LiveVariables should not be necessary here!
484 // PPCTLSDynamicCallPass uses LiveIntervals which previously dependent on
485 // LiveVariables. This (unnecessary) dependency has been removed now,
486 // however a stage-2 clang build fails without LiveVariables computed here.
487 addPass(&LiveVariablesID, false);
488 addPass(createPPCTLSDynamicCallPass());
490 if (EnableExtraTOCRegDeps)
491 addPass(createPPCTOCRegDepsPass());
493 if (getOptLevel() != CodeGenOpt::None)
494 addPass(&MachinePipelinerID);
497 void PPCPassConfig::addPreSched2() {
498 if (getOptLevel() != CodeGenOpt::None) {
499 addPass(&IfConverterID);
501 // This optimization must happen after anything that might do store-to-load
502 // forwarding. Here we're after RA (and, thus, when spills are inserted)
503 // but before post-RA scheduling.
504 if (!DisableQPXLoadSplat)
505 addPass(createPPCQPXLoadSplatPass());
509 void PPCPassConfig::addPreEmitPass() {
510 addPass(createPPCPreEmitPeepholePass());
511 addPass(createPPCExpandISELPass());
513 if (getOptLevel() != CodeGenOpt::None)
514 addPass(createPPCEarlyReturnPass(), false);
515 // Must run branch selection immediately preceding the asm printer.
516 addPass(createPPCBranchSelectionPass(), false);
519 TargetTransformInfo
520 PPCTargetMachine::getTargetTransformInfo(const Function &F) {
521 return TargetTransformInfo(PPCTTIImpl(this, F));
524 static MachineSchedRegistry
525 PPCPreRASchedRegistry("ppc-prera",
526 "Run PowerPC PreRA specific scheduler",
527 createPPCMachineScheduler);
529 static MachineSchedRegistry
530 PPCPostRASchedRegistry("ppc-postra",
531 "Run PowerPC PostRA specific scheduler",
532 createPPCPostMachineScheduler);