[ARM] MVE integer min and max
[llvm-core.git] / lib / Target / PowerPC / PPCTargetMachine.cpp
blobce00f848dd72550c581d19251f8a9d8a577cd2f4
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 it isn't a Mach-O file then it's going to be a linux ELF
189 // object file.
190 if (TT.isOSDarwin())
191 return llvm::make_unique<TargetLoweringObjectFileMachO>();
193 return llvm::make_unique<PPC64LinuxTargetObjectFile>();
196 static PPCTargetMachine::PPCABI computeTargetABI(const Triple &TT,
197 const TargetOptions &Options) {
198 if (TT.isOSDarwin())
199 report_fatal_error("Darwin is no longer supported for PowerPC");
201 if (Options.MCOptions.getABIName().startswith("elfv1"))
202 return PPCTargetMachine::PPC_ABI_ELFv1;
203 else if (Options.MCOptions.getABIName().startswith("elfv2"))
204 return PPCTargetMachine::PPC_ABI_ELFv2;
206 assert(Options.MCOptions.getABIName().empty() &&
207 "Unknown target-abi option!");
209 if (TT.isMacOSX())
210 return PPCTargetMachine::PPC_ABI_UNKNOWN;
212 switch (TT.getArch()) {
213 case Triple::ppc64le:
214 return PPCTargetMachine::PPC_ABI_ELFv2;
215 case Triple::ppc64:
216 if (TT.getEnvironment() == llvm::Triple::ELFv2)
217 return PPCTargetMachine::PPC_ABI_ELFv2;
218 return PPCTargetMachine::PPC_ABI_ELFv1;
219 default:
220 return PPCTargetMachine::PPC_ABI_UNKNOWN;
224 static Reloc::Model getEffectiveRelocModel(const Triple &TT,
225 Optional<Reloc::Model> RM) {
226 if (RM.hasValue())
227 return *RM;
229 // Darwin defaults to dynamic-no-pic.
230 if (TT.isOSDarwin())
231 return Reloc::DynamicNoPIC;
233 // Big Endian PPC is PIC by default.
234 if (TT.getArch() == Triple::ppc64)
235 return Reloc::PIC_;
237 // Rest are static by default.
238 return Reloc::Static;
241 static CodeModel::Model getEffectivePPCCodeModel(const Triple &TT,
242 Optional<CodeModel::Model> CM,
243 bool JIT) {
244 if (CM) {
245 if (*CM == CodeModel::Tiny)
246 report_fatal_error("Target does not support the tiny CodeModel", false);
247 if (*CM == CodeModel::Kernel)
248 report_fatal_error("Target does not support the kernel CodeModel", false);
249 return *CM;
251 if (!TT.isOSDarwin() && !JIT &&
252 (TT.getArch() == Triple::ppc64 || TT.getArch() == Triple::ppc64le))
253 return CodeModel::Medium;
254 return CodeModel::Small;
258 static ScheduleDAGInstrs *createPPCMachineScheduler(MachineSchedContext *C) {
259 const PPCSubtarget &ST = C->MF->getSubtarget<PPCSubtarget>();
260 ScheduleDAGMILive *DAG =
261 new ScheduleDAGMILive(C, ST.usePPCPreRASchedStrategy() ?
262 llvm::make_unique<PPCPreRASchedStrategy>(C) :
263 llvm::make_unique<GenericScheduler>(C));
264 // add DAG Mutations here.
265 DAG->addMutation(createCopyConstrainDAGMutation(DAG->TII, DAG->TRI));
266 return DAG;
269 static ScheduleDAGInstrs *createPPCPostMachineScheduler(
270 MachineSchedContext *C) {
271 const PPCSubtarget &ST = C->MF->getSubtarget<PPCSubtarget>();
272 ScheduleDAGMI *DAG =
273 new ScheduleDAGMI(C, ST.usePPCPostRASchedStrategy() ?
274 llvm::make_unique<PPCPostRASchedStrategy>(C) :
275 llvm::make_unique<PostGenericScheduler>(C), true);
276 // add DAG Mutations here.
277 return DAG;
280 // The FeatureString here is a little subtle. We are modifying the feature
281 // string with what are (currently) non-function specific overrides as it goes
282 // into the LLVMTargetMachine constructor and then using the stored value in the
283 // Subtarget constructor below it.
284 PPCTargetMachine::PPCTargetMachine(const Target &T, const Triple &TT,
285 StringRef CPU, StringRef FS,
286 const TargetOptions &Options,
287 Optional<Reloc::Model> RM,
288 Optional<CodeModel::Model> CM,
289 CodeGenOpt::Level OL, bool JIT)
290 : LLVMTargetMachine(T, getDataLayoutString(TT), TT, CPU,
291 computeFSAdditions(FS, OL, TT), Options,
292 getEffectiveRelocModel(TT, RM),
293 getEffectivePPCCodeModel(TT, CM, JIT), OL),
294 TLOF(createTLOF(getTargetTriple())),
295 TargetABI(computeTargetABI(TT, Options)) {
296 initAsmInfo();
299 PPCTargetMachine::~PPCTargetMachine() = default;
301 const PPCSubtarget *
302 PPCTargetMachine::getSubtargetImpl(const Function &F) const {
303 Attribute CPUAttr = F.getFnAttribute("target-cpu");
304 Attribute FSAttr = F.getFnAttribute("target-features");
306 std::string CPU = !CPUAttr.hasAttribute(Attribute::None)
307 ? CPUAttr.getValueAsString().str()
308 : TargetCPU;
309 std::string FS = !FSAttr.hasAttribute(Attribute::None)
310 ? FSAttr.getValueAsString().str()
311 : TargetFS;
313 // FIXME: This is related to the code below to reset the target options,
314 // we need to know whether or not the soft float flag is set on the
315 // function before we can generate a subtarget. We also need to use
316 // it as a key for the subtarget since that can be the only difference
317 // between two functions.
318 bool SoftFloat =
319 F.getFnAttribute("use-soft-float").getValueAsString() == "true";
320 // If the soft float attribute is set on the function turn on the soft float
321 // subtarget feature.
322 if (SoftFloat)
323 FS += FS.empty() ? "-hard-float" : ",-hard-float";
325 auto &I = SubtargetMap[CPU + FS];
326 if (!I) {
327 // This needs to be done before we create a new subtarget since any
328 // creation will depend on the TM and the code generation flags on the
329 // function that reside in TargetOptions.
330 resetTargetOptions(F);
331 I = llvm::make_unique<PPCSubtarget>(
332 TargetTriple, CPU,
333 // FIXME: It would be good to have the subtarget additions here
334 // not necessary. Anything that turns them on/off (overrides) ends
335 // up being put at the end of the feature string, but the defaults
336 // shouldn't require adding them. Fixing this means pulling Feature64Bit
337 // out of most of the target cpus in the .td file and making it set only
338 // as part of initialization via the TargetTriple.
339 computeFSAdditions(FS, getOptLevel(), getTargetTriple()), *this);
341 return I.get();
344 //===----------------------------------------------------------------------===//
345 // Pass Pipeline Configuration
346 //===----------------------------------------------------------------------===//
348 namespace {
350 /// PPC Code Generator Pass Configuration Options.
351 class PPCPassConfig : public TargetPassConfig {
352 public:
353 PPCPassConfig(PPCTargetMachine &TM, PassManagerBase &PM)
354 : TargetPassConfig(TM, PM) {
355 // At any optimization level above -O0 we use the Machine Scheduler and not
356 // the default Post RA List Scheduler.
357 if (TM.getOptLevel() != CodeGenOpt::None)
358 substitutePass(&PostRASchedulerID, &PostMachineSchedulerID);
361 PPCTargetMachine &getPPCTargetMachine() const {
362 return getTM<PPCTargetMachine>();
365 void addIRPasses() override;
366 bool addPreISel() override;
367 bool addILPOpts() override;
368 bool addInstSelector() override;
369 void addMachineSSAOptimization() override;
370 void addPreRegAlloc() override;
371 void addPreSched2() override;
372 void addPreEmitPass() override;
373 ScheduleDAGInstrs *
374 createMachineScheduler(MachineSchedContext *C) const override {
375 return createPPCMachineScheduler(C);
377 ScheduleDAGInstrs *
378 createPostMachineScheduler(MachineSchedContext *C) const override {
379 return createPPCPostMachineScheduler(C);
383 } // end anonymous namespace
385 TargetPassConfig *PPCTargetMachine::createPassConfig(PassManagerBase &PM) {
386 return new PPCPassConfig(*this, PM);
389 void PPCPassConfig::addIRPasses() {
390 if (TM->getOptLevel() != CodeGenOpt::None)
391 addPass(createPPCBoolRetToIntPass());
392 addPass(createAtomicExpandPass());
394 // For the BG/Q (or if explicitly requested), add explicit data prefetch
395 // intrinsics.
396 bool UsePrefetching = TM->getTargetTriple().getVendor() == Triple::BGQ &&
397 getOptLevel() != CodeGenOpt::None;
398 if (EnablePrefetch.getNumOccurrences() > 0)
399 UsePrefetching = EnablePrefetch;
400 if (UsePrefetching)
401 addPass(createLoopDataPrefetchPass());
403 if (TM->getOptLevel() >= CodeGenOpt::Default && EnableGEPOpt) {
404 // Call SeparateConstOffsetFromGEP pass to extract constants within indices
405 // and lower a GEP with multiple indices to either arithmetic operations or
406 // multiple GEPs with single index.
407 addPass(createSeparateConstOffsetFromGEPPass(true));
408 // Call EarlyCSE pass to find and remove subexpressions in the lowered
409 // result.
410 addPass(createEarlyCSEPass());
411 // Do loop invariant code motion in case part of the lowered result is
412 // invariant.
413 addPass(createLICMPass());
416 TargetPassConfig::addIRPasses();
419 bool PPCPassConfig::addPreISel() {
420 if (!DisablePreIncPrep && getOptLevel() != CodeGenOpt::None)
421 addPass(createPPCLoopPreIncPrepPass(getPPCTargetMachine()));
423 if (!DisableCTRLoops && getOptLevel() != CodeGenOpt::None)
424 addPass(createHardwareLoopsPass());
426 return false;
429 bool PPCPassConfig::addILPOpts() {
430 addPass(&EarlyIfConverterID);
432 if (EnableMachineCombinerPass)
433 addPass(&MachineCombinerID);
435 return true;
438 bool PPCPassConfig::addInstSelector() {
439 // Install an instruction selector.
440 addPass(createPPCISelDag(getPPCTargetMachine(), getOptLevel()));
442 #ifndef NDEBUG
443 if (!DisableCTRLoops && getOptLevel() != CodeGenOpt::None)
444 addPass(createPPCCTRLoopsVerify());
445 #endif
447 addPass(createPPCVSXCopyPass());
448 return false;
451 void PPCPassConfig::addMachineSSAOptimization() {
452 // PPCBranchCoalescingPass need to be done before machine sinking
453 // since it merges empty blocks.
454 if (EnableBranchCoalescing && getOptLevel() != CodeGenOpt::None)
455 addPass(createPPCBranchCoalescingPass());
456 TargetPassConfig::addMachineSSAOptimization();
457 // For little endian, remove where possible the vector swap instructions
458 // introduced at code generation to normalize vector element order.
459 if (TM->getTargetTriple().getArch() == Triple::ppc64le &&
460 !DisableVSXSwapRemoval)
461 addPass(createPPCVSXSwapRemovalPass());
462 // Reduce the number of cr-logical ops.
463 if (ReduceCRLogical && getOptLevel() != CodeGenOpt::None)
464 addPass(createPPCReduceCRLogicalsPass());
465 // Target-specific peephole cleanups performed after instruction
466 // selection.
467 if (!DisableMIPeephole) {
468 addPass(createPPCMIPeepholePass());
469 addPass(&DeadMachineInstructionElimID);
473 void PPCPassConfig::addPreRegAlloc() {
474 if (getOptLevel() != CodeGenOpt::None) {
475 initializePPCVSXFMAMutatePass(*PassRegistry::getPassRegistry());
476 insertPass(VSXFMAMutateEarly ? &RegisterCoalescerID : &MachineSchedulerID,
477 &PPCVSXFMAMutateID);
480 // FIXME: We probably don't need to run these for -fPIE.
481 if (getPPCTargetMachine().isPositionIndependent()) {
482 // FIXME: LiveVariables should not be necessary here!
483 // PPCTLSDynamicCallPass uses LiveIntervals which previously dependent on
484 // LiveVariables. This (unnecessary) dependency has been removed now,
485 // however a stage-2 clang build fails without LiveVariables computed here.
486 addPass(&LiveVariablesID, false);
487 addPass(createPPCTLSDynamicCallPass());
489 if (EnableExtraTOCRegDeps)
490 addPass(createPPCTOCRegDepsPass());
492 if (getOptLevel() != CodeGenOpt::None)
493 addPass(&MachinePipelinerID);
496 void PPCPassConfig::addPreSched2() {
497 if (getOptLevel() != CodeGenOpt::None) {
498 addPass(&IfConverterID);
500 // This optimization must happen after anything that might do store-to-load
501 // forwarding. Here we're after RA (and, thus, when spills are inserted)
502 // but before post-RA scheduling.
503 if (!DisableQPXLoadSplat)
504 addPass(createPPCQPXLoadSplatPass());
508 void PPCPassConfig::addPreEmitPass() {
509 addPass(createPPCPreEmitPeepholePass());
510 addPass(createPPCExpandISELPass());
512 if (getOptLevel() != CodeGenOpt::None)
513 addPass(createPPCEarlyReturnPass(), false);
514 // Must run branch selection immediately preceding the asm printer.
515 addPass(createPPCBranchSelectionPass(), false);
518 TargetTransformInfo
519 PPCTargetMachine::getTargetTransformInfo(const Function &F) {
520 return TargetTransformInfo(PPCTTIImpl(this, F));
523 static MachineSchedRegistry
524 PPCPreRASchedRegistry("ppc-prera",
525 "Run PowerPC PreRA specific scheduler",
526 createPPCMachineScheduler);
528 static MachineSchedRegistry
529 PPCPostRASchedRegistry("ppc-postra",
530 "Run PowerPC PostRA specific scheduler",
531 createPPCPostMachineScheduler);