xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 //===-- AMDGPUCodeGenPrepare.cpp ------------------------------------------===//
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 /// \file
10 /// This pass does misc. AMDGPU optimizations on IR before instruction
11 /// selection.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "AMDGPU.h"
16 #include "AMDGPUTargetMachine.h"
17 #include "SIModeRegisterDefaults.h"
18 #include "llvm/Analysis/AssumptionCache.h"
19 #include "llvm/Analysis/ConstantFolding.h"
20 #include "llvm/Analysis/TargetLibraryInfo.h"
21 #include "llvm/Analysis/UniformityAnalysis.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/CodeGen/TargetPassConfig.h"
24 #include "llvm/IR/Dominators.h"
25 #include "llvm/IR/IRBuilder.h"
26 #include "llvm/IR/InstVisitor.h"
27 #include "llvm/IR/IntrinsicsAMDGPU.h"
28 #include "llvm/IR/PatternMatch.h"
29 #include "llvm/InitializePasses.h"
30 #include "llvm/Pass.h"
31 #include "llvm/Support/KnownBits.h"
32 #include "llvm/Transforms/Utils/IntegerDivision.h"
33 #include "llvm/Transforms/Utils/Local.h"
34 
35 #define DEBUG_TYPE "amdgpu-codegenprepare"
36 
37 using namespace llvm;
38 using namespace llvm::PatternMatch;
39 
40 namespace {
41 
42 static cl::opt<bool> WidenLoads(
43   "amdgpu-codegenprepare-widen-constant-loads",
44   cl::desc("Widen sub-dword constant address space loads in AMDGPUCodeGenPrepare"),
45   cl::ReallyHidden,
46   cl::init(false));
47 
48 static cl::opt<bool> Widen16BitOps(
49   "amdgpu-codegenprepare-widen-16-bit-ops",
50   cl::desc("Widen uniform 16-bit instructions to 32-bit in AMDGPUCodeGenPrepare"),
51   cl::ReallyHidden,
52   cl::init(true));
53 
54 static cl::opt<bool>
55     BreakLargePHIs("amdgpu-codegenprepare-break-large-phis",
56                    cl::desc("Break large PHI nodes for DAGISel"),
57                    cl::ReallyHidden, cl::init(true));
58 
59 static cl::opt<bool>
60     ForceBreakLargePHIs("amdgpu-codegenprepare-force-break-large-phis",
61                         cl::desc("For testing purposes, always break large "
62                                  "PHIs even if it isn't profitable."),
63                         cl::ReallyHidden, cl::init(false));
64 
65 static cl::opt<unsigned> BreakLargePHIsThreshold(
66     "amdgpu-codegenprepare-break-large-phis-threshold",
67     cl::desc("Minimum type size in bits for breaking large PHI nodes"),
68     cl::ReallyHidden, cl::init(32));
69 
70 static cl::opt<bool> UseMul24Intrin(
71   "amdgpu-codegenprepare-mul24",
72   cl::desc("Introduce mul24 intrinsics in AMDGPUCodeGenPrepare"),
73   cl::ReallyHidden,
74   cl::init(true));
75 
76 // Legalize 64-bit division by using the generic IR expansion.
77 static cl::opt<bool> ExpandDiv64InIR(
78   "amdgpu-codegenprepare-expand-div64",
79   cl::desc("Expand 64-bit division in AMDGPUCodeGenPrepare"),
80   cl::ReallyHidden,
81   cl::init(false));
82 
83 // Leave all division operations as they are. This supersedes ExpandDiv64InIR
84 // and is used for testing the legalizer.
85 static cl::opt<bool> DisableIDivExpand(
86   "amdgpu-codegenprepare-disable-idiv-expansion",
87   cl::desc("Prevent expanding integer division in AMDGPUCodeGenPrepare"),
88   cl::ReallyHidden,
89   cl::init(false));
90 
91 // Disable processing of fdiv so we can better test the backend implementations.
92 static cl::opt<bool> DisableFDivExpand(
93   "amdgpu-codegenprepare-disable-fdiv-expansion",
94   cl::desc("Prevent expanding floating point division in AMDGPUCodeGenPrepare"),
95   cl::ReallyHidden,
96   cl::init(false));
97 
98 class AMDGPUCodeGenPrepareImpl
99     : public InstVisitor<AMDGPUCodeGenPrepareImpl, bool> {
100 public:
101   const GCNSubtarget *ST = nullptr;
102   const AMDGPUTargetMachine *TM = nullptr;
103   const TargetLibraryInfo *TLInfo = nullptr;
104   AssumptionCache *AC = nullptr;
105   DominatorTree *DT = nullptr;
106   UniformityInfo *UA = nullptr;
107   Module *Mod = nullptr;
108   const DataLayout *DL = nullptr;
109   bool HasUnsafeFPMath = false;
110   bool HasFP32DenormalFlush = false;
111   bool FlowChanged = false;
112   mutable Function *SqrtF32 = nullptr;
113   mutable Function *LdexpF32 = nullptr;
114 
115   DenseMap<const PHINode *, bool> BreakPhiNodesCache;
116 
getSqrtF32() const117   Function *getSqrtF32() const {
118     if (SqrtF32)
119       return SqrtF32;
120 
121     LLVMContext &Ctx = Mod->getContext();
122     SqrtF32 = Intrinsic::getDeclaration(Mod, Intrinsic::amdgcn_sqrt,
123                                         {Type::getFloatTy(Ctx)});
124     return SqrtF32;
125   }
126 
getLdexpF32() const127   Function *getLdexpF32() const {
128     if (LdexpF32)
129       return LdexpF32;
130 
131     LLVMContext &Ctx = Mod->getContext();
132     LdexpF32 = Intrinsic::getDeclaration(
133         Mod, Intrinsic::ldexp, {Type::getFloatTy(Ctx), Type::getInt32Ty(Ctx)});
134     return LdexpF32;
135   }
136 
137   bool canBreakPHINode(const PHINode &I);
138 
139   /// Copies exact/nsw/nuw flags (if any) from binary operation \p I to
140   /// binary operation \p V.
141   ///
142   /// \returns Binary operation \p V.
143   /// \returns \p T's base element bit width.
144   unsigned getBaseElementBitWidth(const Type *T) const;
145 
146   /// \returns Equivalent 32 bit integer type for given type \p T. For example,
147   /// if \p T is i7, then i32 is returned; if \p T is <3 x i12>, then <3 x i32>
148   /// is returned.
149   Type *getI32Ty(IRBuilder<> &B, const Type *T) const;
150 
151   /// \returns True if binary operation \p I is a signed binary operation, false
152   /// otherwise.
153   bool isSigned(const BinaryOperator &I) const;
154 
155   /// \returns True if the condition of 'select' operation \p I comes from a
156   /// signed 'icmp' operation, false otherwise.
157   bool isSigned(const SelectInst &I) const;
158 
159   /// \returns True if type \p T needs to be promoted to 32 bit integer type,
160   /// false otherwise.
161   bool needsPromotionToI32(const Type *T) const;
162 
163   /// Return true if \p T is a legal scalar floating point type.
164   bool isLegalFloatingTy(const Type *T) const;
165 
166   /// Wrapper to pass all the arguments to computeKnownFPClass
computeKnownFPClass(const Value * V,FPClassTest Interested,const Instruction * CtxI) const167   KnownFPClass computeKnownFPClass(const Value *V, FPClassTest Interested,
168                                    const Instruction *CtxI) const {
169     return llvm::computeKnownFPClass(V, *DL, Interested, 0, TLInfo, AC, CtxI,
170                                      DT);
171   }
172 
canIgnoreDenormalInput(const Value * V,const Instruction * CtxI) const173   bool canIgnoreDenormalInput(const Value *V, const Instruction *CtxI) const {
174     return HasFP32DenormalFlush ||
175            computeKnownFPClass(V, fcSubnormal, CtxI).isKnownNeverSubnormal();
176   }
177 
178   /// Promotes uniform binary operation \p I to equivalent 32 bit binary
179   /// operation.
180   ///
181   /// \details \p I's base element bit width must be greater than 1 and less
182   /// than or equal 16. Promotion is done by sign or zero extending operands to
183   /// 32 bits, replacing \p I with equivalent 32 bit binary operation, and
184   /// truncating the result of 32 bit binary operation back to \p I's original
185   /// type. Division operation is not promoted.
186   ///
187   /// \returns True if \p I is promoted to equivalent 32 bit binary operation,
188   /// false otherwise.
189   bool promoteUniformOpToI32(BinaryOperator &I) const;
190 
191   /// Promotes uniform 'icmp' operation \p I to 32 bit 'icmp' operation.
192   ///
193   /// \details \p I's base element bit width must be greater than 1 and less
194   /// than or equal 16. Promotion is done by sign or zero extending operands to
195   /// 32 bits, and replacing \p I with 32 bit 'icmp' operation.
196   ///
197   /// \returns True.
198   bool promoteUniformOpToI32(ICmpInst &I) const;
199 
200   /// Promotes uniform 'select' operation \p I to 32 bit 'select'
201   /// operation.
202   ///
203   /// \details \p I's base element bit width must be greater than 1 and less
204   /// than or equal 16. Promotion is done by sign or zero extending operands to
205   /// 32 bits, replacing \p I with 32 bit 'select' operation, and truncating the
206   /// result of 32 bit 'select' operation back to \p I's original type.
207   ///
208   /// \returns True.
209   bool promoteUniformOpToI32(SelectInst &I) const;
210 
211   /// Promotes uniform 'bitreverse' intrinsic \p I to 32 bit 'bitreverse'
212   /// intrinsic.
213   ///
214   /// \details \p I's base element bit width must be greater than 1 and less
215   /// than or equal 16. Promotion is done by zero extending the operand to 32
216   /// bits, replacing \p I with 32 bit 'bitreverse' intrinsic, shifting the
217   /// result of 32 bit 'bitreverse' intrinsic to the right with zero fill (the
218   /// shift amount is 32 minus \p I's base element bit width), and truncating
219   /// the result of the shift operation back to \p I's original type.
220   ///
221   /// \returns True.
222   bool promoteUniformBitreverseToI32(IntrinsicInst &I) const;
223 
224   /// \returns The minimum number of bits needed to store the value of \Op as an
225   /// unsigned integer. Truncating to this size and then zero-extending to
226   /// the original will not change the value.
227   unsigned numBitsUnsigned(Value *Op) const;
228 
229   /// \returns The minimum number of bits needed to store the value of \Op as a
230   /// signed integer. Truncating to this size and then sign-extending to
231   /// the original size will not change the value.
232   unsigned numBitsSigned(Value *Op) const;
233 
234   /// Replace mul instructions with llvm.amdgcn.mul.u24 or llvm.amdgcn.mul.s24.
235   /// SelectionDAG has an issue where an and asserting the bits are known
236   bool replaceMulWithMul24(BinaryOperator &I) const;
237 
238   /// Perform same function as equivalently named function in DAGCombiner. Since
239   /// we expand some divisions here, we need to perform this before obscuring.
240   bool foldBinOpIntoSelect(BinaryOperator &I) const;
241 
242   bool divHasSpecialOptimization(BinaryOperator &I,
243                                  Value *Num, Value *Den) const;
244   int getDivNumBits(BinaryOperator &I,
245                     Value *Num, Value *Den,
246                     unsigned AtLeast, bool Signed) const;
247 
248   /// Expands 24 bit div or rem.
249   Value* expandDivRem24(IRBuilder<> &Builder, BinaryOperator &I,
250                         Value *Num, Value *Den,
251                         bool IsDiv, bool IsSigned) const;
252 
253   Value *expandDivRem24Impl(IRBuilder<> &Builder, BinaryOperator &I,
254                             Value *Num, Value *Den, unsigned NumBits,
255                             bool IsDiv, bool IsSigned) const;
256 
257   /// Expands 32 bit div or rem.
258   Value* expandDivRem32(IRBuilder<> &Builder, BinaryOperator &I,
259                         Value *Num, Value *Den) const;
260 
261   Value *shrinkDivRem64(IRBuilder<> &Builder, BinaryOperator &I,
262                         Value *Num, Value *Den) const;
263   void expandDivRem64(BinaryOperator &I) const;
264 
265   /// Widen a scalar load.
266   ///
267   /// \details \p Widen scalar load for uniform, small type loads from constant
268   //  memory / to a full 32-bits and then truncate the input to allow a scalar
269   //  load instead of a vector load.
270   //
271   /// \returns True.
272 
273   bool canWidenScalarExtLoad(LoadInst &I) const;
274 
275   Value *matchFractPat(IntrinsicInst &I);
276   Value *applyFractPat(IRBuilder<> &Builder, Value *FractArg);
277 
278   bool canOptimizeWithRsq(const FPMathOperator *SqrtOp, FastMathFlags DivFMF,
279                           FastMathFlags SqrtFMF) const;
280 
281   Value *optimizeWithRsq(IRBuilder<> &Builder, Value *Num, Value *Den,
282                          FastMathFlags DivFMF, FastMathFlags SqrtFMF,
283                          const Instruction *CtxI) const;
284 
285   Value *optimizeWithRcp(IRBuilder<> &Builder, Value *Num, Value *Den,
286                          FastMathFlags FMF, const Instruction *CtxI) const;
287   Value *optimizeWithFDivFast(IRBuilder<> &Builder, Value *Num, Value *Den,
288                               float ReqdAccuracy) const;
289 
290   Value *visitFDivElement(IRBuilder<> &Builder, Value *Num, Value *Den,
291                           FastMathFlags DivFMF, FastMathFlags SqrtFMF,
292                           Value *RsqOp, const Instruction *FDiv,
293                           float ReqdAccuracy) const;
294 
295   std::pair<Value *, Value *> getFrexpResults(IRBuilder<> &Builder,
296                                               Value *Src) const;
297 
298   Value *emitRcpIEEE1ULP(IRBuilder<> &Builder, Value *Src,
299                          bool IsNegative) const;
300   Value *emitFrexpDiv(IRBuilder<> &Builder, Value *LHS, Value *RHS,
301                       FastMathFlags FMF) const;
302   Value *emitSqrtIEEE2ULP(IRBuilder<> &Builder, Value *Src,
303                           FastMathFlags FMF) const;
304 
305 public:
306   bool visitFDiv(BinaryOperator &I);
307 
visitInstruction(Instruction & I)308   bool visitInstruction(Instruction &I) { return false; }
309   bool visitBinaryOperator(BinaryOperator &I);
310   bool visitLoadInst(LoadInst &I);
311   bool visitICmpInst(ICmpInst &I);
312   bool visitSelectInst(SelectInst &I);
313   bool visitPHINode(PHINode &I);
314   bool visitAddrSpaceCastInst(AddrSpaceCastInst &I);
315 
316   bool visitIntrinsicInst(IntrinsicInst &I);
317   bool visitBitreverseIntrinsicInst(IntrinsicInst &I);
318   bool visitMinNum(IntrinsicInst &I);
319   bool visitSqrt(IntrinsicInst &I);
320   bool run(Function &F);
321 };
322 
323 class AMDGPUCodeGenPrepare : public FunctionPass {
324 private:
325   AMDGPUCodeGenPrepareImpl Impl;
326 
327 public:
328   static char ID;
AMDGPUCodeGenPrepare()329   AMDGPUCodeGenPrepare() : FunctionPass(ID) {
330     initializeAMDGPUCodeGenPreparePass(*PassRegistry::getPassRegistry());
331   }
getAnalysisUsage(AnalysisUsage & AU) const332   void getAnalysisUsage(AnalysisUsage &AU) const override {
333     AU.addRequired<AssumptionCacheTracker>();
334     AU.addRequired<UniformityInfoWrapperPass>();
335     AU.addRequired<TargetLibraryInfoWrapperPass>();
336 
337     // FIXME: Division expansion needs to preserve the dominator tree.
338     if (!ExpandDiv64InIR)
339       AU.setPreservesAll();
340   }
341   bool runOnFunction(Function &F) override;
342   bool doInitialization(Module &M) override;
getPassName() const343   StringRef getPassName() const override { return "AMDGPU IR optimizations"; }
344 };
345 
346 } // end anonymous namespace
347 
run(Function & F)348 bool AMDGPUCodeGenPrepareImpl::run(Function &F) {
349   BreakPhiNodesCache.clear();
350   bool MadeChange = false;
351 
352   Function::iterator NextBB;
353   for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; FI = NextBB) {
354     BasicBlock *BB = &*FI;
355     NextBB = std::next(FI);
356 
357     BasicBlock::iterator Next;
358     for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
359          I = Next) {
360       Next = std::next(I);
361 
362       MadeChange |= visit(*I);
363 
364       if (Next != E) { // Control flow changed
365         BasicBlock *NextInstBB = Next->getParent();
366         if (NextInstBB != BB) {
367           BB = NextInstBB;
368           E = BB->end();
369           FE = F.end();
370         }
371       }
372     }
373   }
374   return MadeChange;
375 }
376 
getBaseElementBitWidth(const Type * T) const377 unsigned AMDGPUCodeGenPrepareImpl::getBaseElementBitWidth(const Type *T) const {
378   assert(needsPromotionToI32(T) && "T does not need promotion to i32");
379 
380   if (T->isIntegerTy())
381     return T->getIntegerBitWidth();
382   return cast<VectorType>(T)->getElementType()->getIntegerBitWidth();
383 }
384 
getI32Ty(IRBuilder<> & B,const Type * T) const385 Type *AMDGPUCodeGenPrepareImpl::getI32Ty(IRBuilder<> &B, const Type *T) const {
386   assert(needsPromotionToI32(T) && "T does not need promotion to i32");
387 
388   if (T->isIntegerTy())
389     return B.getInt32Ty();
390   return FixedVectorType::get(B.getInt32Ty(), cast<FixedVectorType>(T));
391 }
392 
isSigned(const BinaryOperator & I) const393 bool AMDGPUCodeGenPrepareImpl::isSigned(const BinaryOperator &I) const {
394   return I.getOpcode() == Instruction::AShr ||
395       I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::SRem;
396 }
397 
isSigned(const SelectInst & I) const398 bool AMDGPUCodeGenPrepareImpl::isSigned(const SelectInst &I) const {
399   return isa<ICmpInst>(I.getOperand(0)) ?
400       cast<ICmpInst>(I.getOperand(0))->isSigned() : false;
401 }
402 
needsPromotionToI32(const Type * T) const403 bool AMDGPUCodeGenPrepareImpl::needsPromotionToI32(const Type *T) const {
404   if (!Widen16BitOps)
405     return false;
406 
407   const IntegerType *IntTy = dyn_cast<IntegerType>(T);
408   if (IntTy && IntTy->getBitWidth() > 1 && IntTy->getBitWidth() <= 16)
409     return true;
410 
411   if (const VectorType *VT = dyn_cast<VectorType>(T)) {
412     // TODO: The set of packed operations is more limited, so may want to
413     // promote some anyway.
414     if (ST->hasVOP3PInsts())
415       return false;
416 
417     return needsPromotionToI32(VT->getElementType());
418   }
419 
420   return false;
421 }
422 
isLegalFloatingTy(const Type * Ty) const423 bool AMDGPUCodeGenPrepareImpl::isLegalFloatingTy(const Type *Ty) const {
424   return Ty->isFloatTy() || Ty->isDoubleTy() ||
425          (Ty->isHalfTy() && ST->has16BitInsts());
426 }
427 
428 // Return true if the op promoted to i32 should have nsw set.
promotedOpIsNSW(const Instruction & I)429 static bool promotedOpIsNSW(const Instruction &I) {
430   switch (I.getOpcode()) {
431   case Instruction::Shl:
432   case Instruction::Add:
433   case Instruction::Sub:
434     return true;
435   case Instruction::Mul:
436     return I.hasNoUnsignedWrap();
437   default:
438     return false;
439   }
440 }
441 
442 // Return true if the op promoted to i32 should have nuw set.
promotedOpIsNUW(const Instruction & I)443 static bool promotedOpIsNUW(const Instruction &I) {
444   switch (I.getOpcode()) {
445   case Instruction::Shl:
446   case Instruction::Add:
447   case Instruction::Mul:
448     return true;
449   case Instruction::Sub:
450     return I.hasNoUnsignedWrap();
451   default:
452     return false;
453   }
454 }
455 
canWidenScalarExtLoad(LoadInst & I) const456 bool AMDGPUCodeGenPrepareImpl::canWidenScalarExtLoad(LoadInst &I) const {
457   Type *Ty = I.getType();
458   const DataLayout &DL = Mod->getDataLayout();
459   int TySize = DL.getTypeSizeInBits(Ty);
460   Align Alignment = DL.getValueOrABITypeAlignment(I.getAlign(), Ty);
461 
462   return I.isSimple() && TySize < 32 && Alignment >= 4 && UA->isUniform(&I);
463 }
464 
promoteUniformOpToI32(BinaryOperator & I) const465 bool AMDGPUCodeGenPrepareImpl::promoteUniformOpToI32(BinaryOperator &I) const {
466   assert(needsPromotionToI32(I.getType()) &&
467          "I does not need promotion to i32");
468 
469   if (I.getOpcode() == Instruction::SDiv ||
470       I.getOpcode() == Instruction::UDiv ||
471       I.getOpcode() == Instruction::SRem ||
472       I.getOpcode() == Instruction::URem)
473     return false;
474 
475   IRBuilder<> Builder(&I);
476   Builder.SetCurrentDebugLocation(I.getDebugLoc());
477 
478   Type *I32Ty = getI32Ty(Builder, I.getType());
479   Value *ExtOp0 = nullptr;
480   Value *ExtOp1 = nullptr;
481   Value *ExtRes = nullptr;
482   Value *TruncRes = nullptr;
483 
484   if (isSigned(I)) {
485     ExtOp0 = Builder.CreateSExt(I.getOperand(0), I32Ty);
486     ExtOp1 = Builder.CreateSExt(I.getOperand(1), I32Ty);
487   } else {
488     ExtOp0 = Builder.CreateZExt(I.getOperand(0), I32Ty);
489     ExtOp1 = Builder.CreateZExt(I.getOperand(1), I32Ty);
490   }
491 
492   ExtRes = Builder.CreateBinOp(I.getOpcode(), ExtOp0, ExtOp1);
493   if (Instruction *Inst = dyn_cast<Instruction>(ExtRes)) {
494     if (promotedOpIsNSW(cast<Instruction>(I)))
495       Inst->setHasNoSignedWrap();
496 
497     if (promotedOpIsNUW(cast<Instruction>(I)))
498       Inst->setHasNoUnsignedWrap();
499 
500     if (const auto *ExactOp = dyn_cast<PossiblyExactOperator>(&I))
501       Inst->setIsExact(ExactOp->isExact());
502   }
503 
504   TruncRes = Builder.CreateTrunc(ExtRes, I.getType());
505 
506   I.replaceAllUsesWith(TruncRes);
507   I.eraseFromParent();
508 
509   return true;
510 }
511 
promoteUniformOpToI32(ICmpInst & I) const512 bool AMDGPUCodeGenPrepareImpl::promoteUniformOpToI32(ICmpInst &I) const {
513   assert(needsPromotionToI32(I.getOperand(0)->getType()) &&
514          "I does not need promotion to i32");
515 
516   IRBuilder<> Builder(&I);
517   Builder.SetCurrentDebugLocation(I.getDebugLoc());
518 
519   Type *I32Ty = getI32Ty(Builder, I.getOperand(0)->getType());
520   Value *ExtOp0 = nullptr;
521   Value *ExtOp1 = nullptr;
522   Value *NewICmp  = nullptr;
523 
524   if (I.isSigned()) {
525     ExtOp0 = Builder.CreateSExt(I.getOperand(0), I32Ty);
526     ExtOp1 = Builder.CreateSExt(I.getOperand(1), I32Ty);
527   } else {
528     ExtOp0 = Builder.CreateZExt(I.getOperand(0), I32Ty);
529     ExtOp1 = Builder.CreateZExt(I.getOperand(1), I32Ty);
530   }
531   NewICmp = Builder.CreateICmp(I.getPredicate(), ExtOp0, ExtOp1);
532 
533   I.replaceAllUsesWith(NewICmp);
534   I.eraseFromParent();
535 
536   return true;
537 }
538 
promoteUniformOpToI32(SelectInst & I) const539 bool AMDGPUCodeGenPrepareImpl::promoteUniformOpToI32(SelectInst &I) const {
540   assert(needsPromotionToI32(I.getType()) &&
541          "I does not need promotion to i32");
542 
543   IRBuilder<> Builder(&I);
544   Builder.SetCurrentDebugLocation(I.getDebugLoc());
545 
546   Type *I32Ty = getI32Ty(Builder, I.getType());
547   Value *ExtOp1 = nullptr;
548   Value *ExtOp2 = nullptr;
549   Value *ExtRes = nullptr;
550   Value *TruncRes = nullptr;
551 
552   if (isSigned(I)) {
553     ExtOp1 = Builder.CreateSExt(I.getOperand(1), I32Ty);
554     ExtOp2 = Builder.CreateSExt(I.getOperand(2), I32Ty);
555   } else {
556     ExtOp1 = Builder.CreateZExt(I.getOperand(1), I32Ty);
557     ExtOp2 = Builder.CreateZExt(I.getOperand(2), I32Ty);
558   }
559   ExtRes = Builder.CreateSelect(I.getOperand(0), ExtOp1, ExtOp2);
560   TruncRes = Builder.CreateTrunc(ExtRes, I.getType());
561 
562   I.replaceAllUsesWith(TruncRes);
563   I.eraseFromParent();
564 
565   return true;
566 }
567 
promoteUniformBitreverseToI32(IntrinsicInst & I) const568 bool AMDGPUCodeGenPrepareImpl::promoteUniformBitreverseToI32(
569     IntrinsicInst &I) const {
570   assert(I.getIntrinsicID() == Intrinsic::bitreverse &&
571          "I must be bitreverse intrinsic");
572   assert(needsPromotionToI32(I.getType()) &&
573          "I does not need promotion to i32");
574 
575   IRBuilder<> Builder(&I);
576   Builder.SetCurrentDebugLocation(I.getDebugLoc());
577 
578   Type *I32Ty = getI32Ty(Builder, I.getType());
579   Function *I32 =
580       Intrinsic::getDeclaration(Mod, Intrinsic::bitreverse, { I32Ty });
581   Value *ExtOp = Builder.CreateZExt(I.getOperand(0), I32Ty);
582   Value *ExtRes = Builder.CreateCall(I32, { ExtOp });
583   Value *LShrOp =
584       Builder.CreateLShr(ExtRes, 32 - getBaseElementBitWidth(I.getType()));
585   Value *TruncRes =
586       Builder.CreateTrunc(LShrOp, I.getType());
587 
588   I.replaceAllUsesWith(TruncRes);
589   I.eraseFromParent();
590 
591   return true;
592 }
593 
numBitsUnsigned(Value * Op) const594 unsigned AMDGPUCodeGenPrepareImpl::numBitsUnsigned(Value *Op) const {
595   return computeKnownBits(Op, *DL, 0, AC).countMaxActiveBits();
596 }
597 
numBitsSigned(Value * Op) const598 unsigned AMDGPUCodeGenPrepareImpl::numBitsSigned(Value *Op) const {
599   return ComputeMaxSignificantBits(Op, *DL, 0, AC);
600 }
601 
extractValues(IRBuilder<> & Builder,SmallVectorImpl<Value * > & Values,Value * V)602 static void extractValues(IRBuilder<> &Builder,
603                           SmallVectorImpl<Value *> &Values, Value *V) {
604   auto *VT = dyn_cast<FixedVectorType>(V->getType());
605   if (!VT) {
606     Values.push_back(V);
607     return;
608   }
609 
610   for (int I = 0, E = VT->getNumElements(); I != E; ++I)
611     Values.push_back(Builder.CreateExtractElement(V, I));
612 }
613 
insertValues(IRBuilder<> & Builder,Type * Ty,SmallVectorImpl<Value * > & Values)614 static Value *insertValues(IRBuilder<> &Builder,
615                            Type *Ty,
616                            SmallVectorImpl<Value *> &Values) {
617   if (!Ty->isVectorTy()) {
618     assert(Values.size() == 1);
619     return Values[0];
620   }
621 
622   Value *NewVal = PoisonValue::get(Ty);
623   for (int I = 0, E = Values.size(); I != E; ++I)
624     NewVal = Builder.CreateInsertElement(NewVal, Values[I], I);
625 
626   return NewVal;
627 }
628 
replaceMulWithMul24(BinaryOperator & I) const629 bool AMDGPUCodeGenPrepareImpl::replaceMulWithMul24(BinaryOperator &I) const {
630   if (I.getOpcode() != Instruction::Mul)
631     return false;
632 
633   Type *Ty = I.getType();
634   unsigned Size = Ty->getScalarSizeInBits();
635   if (Size <= 16 && ST->has16BitInsts())
636     return false;
637 
638   // Prefer scalar if this could be s_mul_i32
639   if (UA->isUniform(&I))
640     return false;
641 
642   Value *LHS = I.getOperand(0);
643   Value *RHS = I.getOperand(1);
644   IRBuilder<> Builder(&I);
645   Builder.SetCurrentDebugLocation(I.getDebugLoc());
646 
647   unsigned LHSBits = 0, RHSBits = 0;
648   bool IsSigned = false;
649 
650   if (ST->hasMulU24() && (LHSBits = numBitsUnsigned(LHS)) <= 24 &&
651       (RHSBits = numBitsUnsigned(RHS)) <= 24) {
652     IsSigned = false;
653 
654   } else if (ST->hasMulI24() && (LHSBits = numBitsSigned(LHS)) <= 24 &&
655              (RHSBits = numBitsSigned(RHS)) <= 24) {
656     IsSigned = true;
657 
658   } else
659     return false;
660 
661   SmallVector<Value *, 4> LHSVals;
662   SmallVector<Value *, 4> RHSVals;
663   SmallVector<Value *, 4> ResultVals;
664   extractValues(Builder, LHSVals, LHS);
665   extractValues(Builder, RHSVals, RHS);
666 
667   IntegerType *I32Ty = Builder.getInt32Ty();
668   IntegerType *IntrinTy = Size > 32 ? Builder.getInt64Ty() : I32Ty;
669   Type *DstTy = LHSVals[0]->getType();
670 
671   for (int I = 0, E = LHSVals.size(); I != E; ++I) {
672     Value *LHS = IsSigned ? Builder.CreateSExtOrTrunc(LHSVals[I], I32Ty)
673                           : Builder.CreateZExtOrTrunc(LHSVals[I], I32Ty);
674     Value *RHS = IsSigned ? Builder.CreateSExtOrTrunc(RHSVals[I], I32Ty)
675                           : Builder.CreateZExtOrTrunc(RHSVals[I], I32Ty);
676     Intrinsic::ID ID =
677         IsSigned ? Intrinsic::amdgcn_mul_i24 : Intrinsic::amdgcn_mul_u24;
678     Value *Result = Builder.CreateIntrinsic(ID, {IntrinTy}, {LHS, RHS});
679     Result = IsSigned ? Builder.CreateSExtOrTrunc(Result, DstTy)
680                       : Builder.CreateZExtOrTrunc(Result, DstTy);
681     ResultVals.push_back(Result);
682   }
683 
684   Value *NewVal = insertValues(Builder, Ty, ResultVals);
685   NewVal->takeName(&I);
686   I.replaceAllUsesWith(NewVal);
687   I.eraseFromParent();
688 
689   return true;
690 }
691 
692 // Find a select instruction, which may have been casted. This is mostly to deal
693 // with cases where i16 selects were promoted here to i32.
findSelectThroughCast(Value * V,CastInst * & Cast)694 static SelectInst *findSelectThroughCast(Value *V, CastInst *&Cast) {
695   Cast = nullptr;
696   if (SelectInst *Sel = dyn_cast<SelectInst>(V))
697     return Sel;
698 
699   if ((Cast = dyn_cast<CastInst>(V))) {
700     if (SelectInst *Sel = dyn_cast<SelectInst>(Cast->getOperand(0)))
701       return Sel;
702   }
703 
704   return nullptr;
705 }
706 
foldBinOpIntoSelect(BinaryOperator & BO) const707 bool AMDGPUCodeGenPrepareImpl::foldBinOpIntoSelect(BinaryOperator &BO) const {
708   // Don't do this unless the old select is going away. We want to eliminate the
709   // binary operator, not replace a binop with a select.
710   int SelOpNo = 0;
711 
712   CastInst *CastOp;
713 
714   // TODO: Should probably try to handle some cases with multiple
715   // users. Duplicating the select may be profitable for division.
716   SelectInst *Sel = findSelectThroughCast(BO.getOperand(0), CastOp);
717   if (!Sel || !Sel->hasOneUse()) {
718     SelOpNo = 1;
719     Sel = findSelectThroughCast(BO.getOperand(1), CastOp);
720   }
721 
722   if (!Sel || !Sel->hasOneUse())
723     return false;
724 
725   Constant *CT = dyn_cast<Constant>(Sel->getTrueValue());
726   Constant *CF = dyn_cast<Constant>(Sel->getFalseValue());
727   Constant *CBO = dyn_cast<Constant>(BO.getOperand(SelOpNo ^ 1));
728   if (!CBO || !CT || !CF)
729     return false;
730 
731   if (CastOp) {
732     if (!CastOp->hasOneUse())
733       return false;
734     CT = ConstantFoldCastOperand(CastOp->getOpcode(), CT, BO.getType(), *DL);
735     CF = ConstantFoldCastOperand(CastOp->getOpcode(), CF, BO.getType(), *DL);
736   }
737 
738   // TODO: Handle special 0/-1 cases DAG combine does, although we only really
739   // need to handle divisions here.
740   Constant *FoldedT = SelOpNo ?
741     ConstantFoldBinaryOpOperands(BO.getOpcode(), CBO, CT, *DL) :
742     ConstantFoldBinaryOpOperands(BO.getOpcode(), CT, CBO, *DL);
743   if (!FoldedT || isa<ConstantExpr>(FoldedT))
744     return false;
745 
746   Constant *FoldedF = SelOpNo ?
747     ConstantFoldBinaryOpOperands(BO.getOpcode(), CBO, CF, *DL) :
748     ConstantFoldBinaryOpOperands(BO.getOpcode(), CF, CBO, *DL);
749   if (!FoldedF || isa<ConstantExpr>(FoldedF))
750     return false;
751 
752   IRBuilder<> Builder(&BO);
753   Builder.SetCurrentDebugLocation(BO.getDebugLoc());
754   if (const FPMathOperator *FPOp = dyn_cast<const FPMathOperator>(&BO))
755     Builder.setFastMathFlags(FPOp->getFastMathFlags());
756 
757   Value *NewSelect = Builder.CreateSelect(Sel->getCondition(),
758                                           FoldedT, FoldedF);
759   NewSelect->takeName(&BO);
760   BO.replaceAllUsesWith(NewSelect);
761   BO.eraseFromParent();
762   if (CastOp)
763     CastOp->eraseFromParent();
764   Sel->eraseFromParent();
765   return true;
766 }
767 
768 std::pair<Value *, Value *>
getFrexpResults(IRBuilder<> & Builder,Value * Src) const769 AMDGPUCodeGenPrepareImpl::getFrexpResults(IRBuilder<> &Builder,
770                                           Value *Src) const {
771   Type *Ty = Src->getType();
772   Value *Frexp = Builder.CreateIntrinsic(Intrinsic::frexp,
773                                          {Ty, Builder.getInt32Ty()}, Src);
774   Value *FrexpMant = Builder.CreateExtractValue(Frexp, {0});
775 
776   // Bypass the bug workaround for the exponent result since it doesn't matter.
777   // TODO: Does the bug workaround even really need to consider the exponent
778   // result? It's unspecified by the spec.
779 
780   Value *FrexpExp =
781       ST->hasFractBug()
782           ? Builder.CreateIntrinsic(Intrinsic::amdgcn_frexp_exp,
783                                     {Builder.getInt32Ty(), Ty}, Src)
784           : Builder.CreateExtractValue(Frexp, {1});
785   return {FrexpMant, FrexpExp};
786 }
787 
788 /// Emit an expansion of 1.0 / Src good for 1ulp that supports denormals.
emitRcpIEEE1ULP(IRBuilder<> & Builder,Value * Src,bool IsNegative) const789 Value *AMDGPUCodeGenPrepareImpl::emitRcpIEEE1ULP(IRBuilder<> &Builder,
790                                                  Value *Src,
791                                                  bool IsNegative) const {
792   // Same as for 1.0, but expand the sign out of the constant.
793   // -1.0 / x -> rcp (fneg x)
794   if (IsNegative)
795     Src = Builder.CreateFNeg(Src);
796 
797   // The rcp instruction doesn't support denormals, so scale the input
798   // out of the denormal range and convert at the end.
799   //
800   // Expand as 2^-n * (1.0 / (x * 2^n))
801 
802   // TODO: Skip scaling if input is known never denormal and the input
803   // range won't underflow to denormal. The hard part is knowing the
804   // result. We need a range check, the result could be denormal for
805   // 0x1p+126 < den <= 0x1p+127.
806   auto [FrexpMant, FrexpExp] = getFrexpResults(Builder, Src);
807   Value *ScaleFactor = Builder.CreateNeg(FrexpExp);
808   Value *Rcp = Builder.CreateUnaryIntrinsic(Intrinsic::amdgcn_rcp, FrexpMant);
809   return Builder.CreateCall(getLdexpF32(), {Rcp, ScaleFactor});
810 }
811 
812 /// Emit a 2ulp expansion for fdiv by using frexp for input scaling.
emitFrexpDiv(IRBuilder<> & Builder,Value * LHS,Value * RHS,FastMathFlags FMF) const813 Value *AMDGPUCodeGenPrepareImpl::emitFrexpDiv(IRBuilder<> &Builder, Value *LHS,
814                                               Value *RHS,
815                                               FastMathFlags FMF) const {
816   // If we have have to work around the fract/frexp bug, we're worse off than
817   // using the fdiv.fast expansion. The full safe expansion is faster if we have
818   // fast FMA.
819   if (HasFP32DenormalFlush && ST->hasFractBug() && !ST->hasFastFMAF32() &&
820       (!FMF.noNaNs() || !FMF.noInfs()))
821     return nullptr;
822 
823   // We're scaling the LHS to avoid a denormal input, and scale the denominator
824   // to avoid large values underflowing the result.
825   auto [FrexpMantRHS, FrexpExpRHS] = getFrexpResults(Builder, RHS);
826 
827   Value *Rcp =
828       Builder.CreateUnaryIntrinsic(Intrinsic::amdgcn_rcp, FrexpMantRHS);
829 
830   auto [FrexpMantLHS, FrexpExpLHS] = getFrexpResults(Builder, LHS);
831   Value *Mul = Builder.CreateFMul(FrexpMantLHS, Rcp);
832 
833   // We multiplied by 2^N/2^M, so we need to multiply by 2^(N-M) to scale the
834   // result.
835   Value *ExpDiff = Builder.CreateSub(FrexpExpLHS, FrexpExpRHS);
836   return Builder.CreateCall(getLdexpF32(), {Mul, ExpDiff});
837 }
838 
839 /// Emit a sqrt that handles denormals and is accurate to 2ulp.
emitSqrtIEEE2ULP(IRBuilder<> & Builder,Value * Src,FastMathFlags FMF) const840 Value *AMDGPUCodeGenPrepareImpl::emitSqrtIEEE2ULP(IRBuilder<> &Builder,
841                                                   Value *Src,
842                                                   FastMathFlags FMF) const {
843   Type *Ty = Src->getType();
844   APFloat SmallestNormal =
845       APFloat::getSmallestNormalized(Ty->getFltSemantics());
846   Value *NeedScale =
847       Builder.CreateFCmpOLT(Src, ConstantFP::get(Ty, SmallestNormal));
848 
849   ConstantInt *Zero = Builder.getInt32(0);
850   Value *InputScaleFactor =
851       Builder.CreateSelect(NeedScale, Builder.getInt32(32), Zero);
852 
853   Value *Scaled = Builder.CreateCall(getLdexpF32(), {Src, InputScaleFactor});
854 
855   Value *Sqrt = Builder.CreateCall(getSqrtF32(), Scaled);
856 
857   Value *OutputScaleFactor =
858       Builder.CreateSelect(NeedScale, Builder.getInt32(-16), Zero);
859   return Builder.CreateCall(getLdexpF32(), {Sqrt, OutputScaleFactor});
860 }
861 
862 /// Emit an expansion of 1.0 / sqrt(Src) good for 1ulp that supports denormals.
emitRsqIEEE1ULP(IRBuilder<> & Builder,Value * Src,bool IsNegative)863 static Value *emitRsqIEEE1ULP(IRBuilder<> &Builder, Value *Src,
864                               bool IsNegative) {
865   // bool need_scale = x < 0x1p-126f;
866   // float input_scale = need_scale ? 0x1.0p+24f : 1.0f;
867   // float output_scale = need_scale ? 0x1.0p+12f : 1.0f;
868   // rsq(x * input_scale) * output_scale;
869 
870   Type *Ty = Src->getType();
871   APFloat SmallestNormal =
872       APFloat::getSmallestNormalized(Ty->getFltSemantics());
873   Value *NeedScale =
874       Builder.CreateFCmpOLT(Src, ConstantFP::get(Ty, SmallestNormal));
875   Constant *One = ConstantFP::get(Ty, 1.0);
876   Constant *InputScale = ConstantFP::get(Ty, 0x1.0p+24);
877   Constant *OutputScale =
878       ConstantFP::get(Ty, IsNegative ? -0x1.0p+12 : 0x1.0p+12);
879 
880   Value *InputScaleFactor = Builder.CreateSelect(NeedScale, InputScale, One);
881 
882   Value *ScaledInput = Builder.CreateFMul(Src, InputScaleFactor);
883   Value *Rsq = Builder.CreateUnaryIntrinsic(Intrinsic::amdgcn_rsq, ScaledInput);
884   Value *OutputScaleFactor = Builder.CreateSelect(
885       NeedScale, OutputScale, IsNegative ? ConstantFP::get(Ty, -1.0) : One);
886 
887   return Builder.CreateFMul(Rsq, OutputScaleFactor);
888 }
889 
canOptimizeWithRsq(const FPMathOperator * SqrtOp,FastMathFlags DivFMF,FastMathFlags SqrtFMF) const890 bool AMDGPUCodeGenPrepareImpl::canOptimizeWithRsq(const FPMathOperator *SqrtOp,
891                                                   FastMathFlags DivFMF,
892                                                   FastMathFlags SqrtFMF) const {
893   // The rsqrt contraction increases accuracy from ~2ulp to ~1ulp.
894   if (!DivFMF.allowContract() || !SqrtFMF.allowContract())
895     return false;
896 
897   // v_rsq_f32 gives 1ulp
898   return SqrtFMF.approxFunc() || HasUnsafeFPMath ||
899          SqrtOp->getFPAccuracy() >= 1.0f;
900 }
901 
optimizeWithRsq(IRBuilder<> & Builder,Value * Num,Value * Den,const FastMathFlags DivFMF,const FastMathFlags SqrtFMF,const Instruction * CtxI) const902 Value *AMDGPUCodeGenPrepareImpl::optimizeWithRsq(
903     IRBuilder<> &Builder, Value *Num, Value *Den, const FastMathFlags DivFMF,
904     const FastMathFlags SqrtFMF, const Instruction *CtxI) const {
905   // The rsqrt contraction increases accuracy from ~2ulp to ~1ulp.
906   assert(DivFMF.allowContract() && SqrtFMF.allowContract());
907 
908   // rsq_f16 is accurate to 0.51 ulp.
909   // rsq_f32 is accurate for !fpmath >= 1.0ulp and denormals are flushed.
910   // rsq_f64 is never accurate.
911   const ConstantFP *CLHS = dyn_cast<ConstantFP>(Num);
912   if (!CLHS)
913     return nullptr;
914 
915   assert(Den->getType()->isFloatTy());
916 
917   bool IsNegative = false;
918 
919   // TODO: Handle other numerator values with arcp.
920   if (CLHS->isExactlyValue(1.0) || (IsNegative = CLHS->isExactlyValue(-1.0))) {
921     // Add in the sqrt flags.
922     IRBuilder<>::FastMathFlagGuard Guard(Builder);
923     Builder.setFastMathFlags(DivFMF | SqrtFMF);
924 
925     if ((DivFMF.approxFunc() && SqrtFMF.approxFunc()) || HasUnsafeFPMath ||
926         canIgnoreDenormalInput(Den, CtxI)) {
927       Value *Result = Builder.CreateUnaryIntrinsic(Intrinsic::amdgcn_rsq, Den);
928       // -1.0 / sqrt(x) -> fneg(rsq(x))
929       return IsNegative ? Builder.CreateFNeg(Result) : Result;
930     }
931 
932     return emitRsqIEEE1ULP(Builder, Den, IsNegative);
933   }
934 
935   return nullptr;
936 }
937 
938 // Optimize fdiv with rcp:
939 //
940 // 1/x -> rcp(x) when rcp is sufficiently accurate or inaccurate rcp is
941 //               allowed with unsafe-fp-math or afn.
942 //
943 // a/b -> a*rcp(b) when arcp is allowed, and we only need provide ULP 1.0
944 Value *
optimizeWithRcp(IRBuilder<> & Builder,Value * Num,Value * Den,FastMathFlags FMF,const Instruction * CtxI) const945 AMDGPUCodeGenPrepareImpl::optimizeWithRcp(IRBuilder<> &Builder, Value *Num,
946                                           Value *Den, FastMathFlags FMF,
947                                           const Instruction *CtxI) const {
948   // rcp_f16 is accurate to 0.51 ulp.
949   // rcp_f32 is accurate for !fpmath >= 1.0ulp and denormals are flushed.
950   // rcp_f64 is never accurate.
951   assert(Den->getType()->isFloatTy());
952 
953   if (const ConstantFP *CLHS = dyn_cast<ConstantFP>(Num)) {
954     bool IsNegative = false;
955     if (CLHS->isExactlyValue(1.0) ||
956         (IsNegative = CLHS->isExactlyValue(-1.0))) {
957       Value *Src = Den;
958 
959       if (HasFP32DenormalFlush || FMF.approxFunc()) {
960         // -1.0 / x -> 1.0 / fneg(x)
961         if (IsNegative)
962           Src = Builder.CreateFNeg(Src);
963 
964         // v_rcp_f32 and v_rsq_f32 do not support denormals, and according to
965         // the CI documentation has a worst case error of 1 ulp.
966         // OpenCL requires <= 2.5 ulp for 1.0 / x, so it should always be OK
967         // to use it as long as we aren't trying to use denormals.
968         //
969         // v_rcp_f16 and v_rsq_f16 DO support denormals.
970 
971         // NOTE: v_sqrt and v_rcp will be combined to v_rsq later. So we don't
972         //       insert rsq intrinsic here.
973 
974         // 1.0 / x -> rcp(x)
975         return Builder.CreateUnaryIntrinsic(Intrinsic::amdgcn_rcp, Src);
976       }
977 
978       // TODO: If the input isn't denormal, and we know the input exponent isn't
979       // big enough to introduce a denormal we can avoid the scaling.
980       return emitRcpIEEE1ULP(Builder, Src, IsNegative);
981     }
982   }
983 
984   if (FMF.allowReciprocal()) {
985     // x / y -> x * (1.0 / y)
986 
987     // TODO: Could avoid denormal scaling and use raw rcp if we knew the output
988     // will never underflow.
989     if (HasFP32DenormalFlush || FMF.approxFunc()) {
990       Value *Recip = Builder.CreateUnaryIntrinsic(Intrinsic::amdgcn_rcp, Den);
991       return Builder.CreateFMul(Num, Recip);
992     }
993 
994     Value *Recip = emitRcpIEEE1ULP(Builder, Den, false);
995     return Builder.CreateFMul(Num, Recip);
996   }
997 
998   return nullptr;
999 }
1000 
1001 // optimize with fdiv.fast:
1002 //
1003 // a/b -> fdiv.fast(a, b) when !fpmath >= 2.5ulp with denormals flushed.
1004 //
1005 // 1/x -> fdiv.fast(1,x)  when !fpmath >= 2.5ulp.
1006 //
1007 // NOTE: optimizeWithRcp should be tried first because rcp is the preference.
optimizeWithFDivFast(IRBuilder<> & Builder,Value * Num,Value * Den,float ReqdAccuracy) const1008 Value *AMDGPUCodeGenPrepareImpl::optimizeWithFDivFast(
1009     IRBuilder<> &Builder, Value *Num, Value *Den, float ReqdAccuracy) const {
1010   // fdiv.fast can achieve 2.5 ULP accuracy.
1011   if (ReqdAccuracy < 2.5f)
1012     return nullptr;
1013 
1014   // Only have fdiv.fast for f32.
1015   assert(Den->getType()->isFloatTy());
1016 
1017   bool NumIsOne = false;
1018   if (const ConstantFP *CNum = dyn_cast<ConstantFP>(Num)) {
1019     if (CNum->isExactlyValue(+1.0) || CNum->isExactlyValue(-1.0))
1020       NumIsOne = true;
1021   }
1022 
1023   // fdiv does not support denormals. But 1.0/x is always fine to use it.
1024   //
1025   // TODO: This works for any value with a specific known exponent range, don't
1026   // just limit to constant 1.
1027   if (!HasFP32DenormalFlush && !NumIsOne)
1028     return nullptr;
1029 
1030   return Builder.CreateIntrinsic(Intrinsic::amdgcn_fdiv_fast, {}, {Num, Den});
1031 }
1032 
visitFDivElement(IRBuilder<> & Builder,Value * Num,Value * Den,FastMathFlags DivFMF,FastMathFlags SqrtFMF,Value * RsqOp,const Instruction * FDivInst,float ReqdDivAccuracy) const1033 Value *AMDGPUCodeGenPrepareImpl::visitFDivElement(
1034     IRBuilder<> &Builder, Value *Num, Value *Den, FastMathFlags DivFMF,
1035     FastMathFlags SqrtFMF, Value *RsqOp, const Instruction *FDivInst,
1036     float ReqdDivAccuracy) const {
1037   if (RsqOp) {
1038     Value *Rsq =
1039         optimizeWithRsq(Builder, Num, RsqOp, DivFMF, SqrtFMF, FDivInst);
1040     if (Rsq)
1041       return Rsq;
1042   }
1043 
1044   Value *Rcp = optimizeWithRcp(Builder, Num, Den, DivFMF, FDivInst);
1045   if (Rcp)
1046     return Rcp;
1047 
1048   // In the basic case fdiv_fast has the same instruction count as the frexp div
1049   // expansion. Slightly prefer fdiv_fast since it ends in an fmul that can
1050   // potentially be fused into a user. Also, materialization of the constants
1051   // can be reused for multiple instances.
1052   Value *FDivFast = optimizeWithFDivFast(Builder, Num, Den, ReqdDivAccuracy);
1053   if (FDivFast)
1054     return FDivFast;
1055 
1056   return emitFrexpDiv(Builder, Num, Den, DivFMF);
1057 }
1058 
1059 // Optimizations is performed based on fpmath, fast math flags as well as
1060 // denormals to optimize fdiv with either rcp or fdiv.fast.
1061 //
1062 // With rcp:
1063 //   1/x -> rcp(x) when rcp is sufficiently accurate or inaccurate rcp is
1064 //                 allowed with unsafe-fp-math or afn.
1065 //
1066 //   a/b -> a*rcp(b) when inaccurate rcp is allowed with unsafe-fp-math or afn.
1067 //
1068 // With fdiv.fast:
1069 //   a/b -> fdiv.fast(a, b) when !fpmath >= 2.5ulp with denormals flushed.
1070 //
1071 //   1/x -> fdiv.fast(1,x)  when !fpmath >= 2.5ulp.
1072 //
1073 // NOTE: rcp is the preference in cases that both are legal.
visitFDiv(BinaryOperator & FDiv)1074 bool AMDGPUCodeGenPrepareImpl::visitFDiv(BinaryOperator &FDiv) {
1075   if (DisableFDivExpand)
1076     return false;
1077 
1078   Type *Ty = FDiv.getType()->getScalarType();
1079   if (!Ty->isFloatTy())
1080     return false;
1081 
1082   // The f64 rcp/rsq approximations are pretty inaccurate. We can do an
1083   // expansion around them in codegen. f16 is good enough to always use.
1084 
1085   const FPMathOperator *FPOp = cast<const FPMathOperator>(&FDiv);
1086   const FastMathFlags DivFMF = FPOp->getFastMathFlags();
1087   const float ReqdAccuracy = FPOp->getFPAccuracy();
1088 
1089   FastMathFlags SqrtFMF;
1090 
1091   Value *Num = FDiv.getOperand(0);
1092   Value *Den = FDiv.getOperand(1);
1093 
1094   Value *RsqOp = nullptr;
1095   auto *DenII = dyn_cast<IntrinsicInst>(Den);
1096   if (DenII && DenII->getIntrinsicID() == Intrinsic::sqrt &&
1097       DenII->hasOneUse()) {
1098     const auto *SqrtOp = cast<FPMathOperator>(DenII);
1099     SqrtFMF = SqrtOp->getFastMathFlags();
1100     if (canOptimizeWithRsq(SqrtOp, DivFMF, SqrtFMF))
1101       RsqOp = SqrtOp->getOperand(0);
1102   }
1103 
1104   // Inaccurate rcp is allowed with unsafe-fp-math or afn.
1105   //
1106   // Defer to codegen to handle this.
1107   //
1108   // TODO: Decide on an interpretation for interactions between afn + arcp +
1109   // !fpmath, and make it consistent between here and codegen. For now, defer
1110   // expansion of afn to codegen. The current interpretation is so aggressive we
1111   // don't need any pre-consideration here when we have better information. A
1112   // more conservative interpretation could use handling here.
1113   const bool AllowInaccurateRcp = HasUnsafeFPMath || DivFMF.approxFunc();
1114   if (!RsqOp && AllowInaccurateRcp)
1115     return false;
1116 
1117   // Defer the correct implementations to codegen.
1118   if (ReqdAccuracy < 1.0f)
1119     return false;
1120 
1121   IRBuilder<> Builder(FDiv.getParent(), std::next(FDiv.getIterator()));
1122   Builder.setFastMathFlags(DivFMF);
1123   Builder.SetCurrentDebugLocation(FDiv.getDebugLoc());
1124 
1125   SmallVector<Value *, 4> NumVals;
1126   SmallVector<Value *, 4> DenVals;
1127   SmallVector<Value *, 4> RsqDenVals;
1128   extractValues(Builder, NumVals, Num);
1129   extractValues(Builder, DenVals, Den);
1130 
1131   if (RsqOp)
1132     extractValues(Builder, RsqDenVals, RsqOp);
1133 
1134   SmallVector<Value *, 4> ResultVals(NumVals.size());
1135   for (int I = 0, E = NumVals.size(); I != E; ++I) {
1136     Value *NumElt = NumVals[I];
1137     Value *DenElt = DenVals[I];
1138     Value *RsqDenElt = RsqOp ? RsqDenVals[I] : nullptr;
1139 
1140     Value *NewElt =
1141         visitFDivElement(Builder, NumElt, DenElt, DivFMF, SqrtFMF, RsqDenElt,
1142                          cast<Instruction>(FPOp), ReqdAccuracy);
1143     if (!NewElt) {
1144       // Keep the original, but scalarized.
1145 
1146       // This has the unfortunate side effect of sometimes scalarizing when
1147       // we're not going to do anything.
1148       NewElt = Builder.CreateFDiv(NumElt, DenElt);
1149       if (auto *NewEltInst = dyn_cast<Instruction>(NewElt))
1150         NewEltInst->copyMetadata(FDiv);
1151     }
1152 
1153     ResultVals[I] = NewElt;
1154   }
1155 
1156   Value *NewVal = insertValues(Builder, FDiv.getType(), ResultVals);
1157 
1158   if (NewVal) {
1159     FDiv.replaceAllUsesWith(NewVal);
1160     NewVal->takeName(&FDiv);
1161     RecursivelyDeleteTriviallyDeadInstructions(&FDiv, TLInfo);
1162   }
1163 
1164   return true;
1165 }
1166 
hasUnsafeFPMath(const Function & F)1167 static bool hasUnsafeFPMath(const Function &F) {
1168   Attribute Attr = F.getFnAttribute("unsafe-fp-math");
1169   return Attr.getValueAsBool();
1170 }
1171 
getMul64(IRBuilder<> & Builder,Value * LHS,Value * RHS)1172 static std::pair<Value*, Value*> getMul64(IRBuilder<> &Builder,
1173                                           Value *LHS, Value *RHS) {
1174   Type *I32Ty = Builder.getInt32Ty();
1175   Type *I64Ty = Builder.getInt64Ty();
1176 
1177   Value *LHS_EXT64 = Builder.CreateZExt(LHS, I64Ty);
1178   Value *RHS_EXT64 = Builder.CreateZExt(RHS, I64Ty);
1179   Value *MUL64 = Builder.CreateMul(LHS_EXT64, RHS_EXT64);
1180   Value *Lo = Builder.CreateTrunc(MUL64, I32Ty);
1181   Value *Hi = Builder.CreateLShr(MUL64, Builder.getInt64(32));
1182   Hi = Builder.CreateTrunc(Hi, I32Ty);
1183   return std::pair(Lo, Hi);
1184 }
1185 
getMulHu(IRBuilder<> & Builder,Value * LHS,Value * RHS)1186 static Value* getMulHu(IRBuilder<> &Builder, Value *LHS, Value *RHS) {
1187   return getMul64(Builder, LHS, RHS).second;
1188 }
1189 
1190 /// Figure out how many bits are really needed for this division. \p AtLeast is
1191 /// an optimization hint to bypass the second ComputeNumSignBits call if we the
1192 /// first one is insufficient. Returns -1 on failure.
getDivNumBits(BinaryOperator & I,Value * Num,Value * Den,unsigned AtLeast,bool IsSigned) const1193 int AMDGPUCodeGenPrepareImpl::getDivNumBits(BinaryOperator &I, Value *Num,
1194                                             Value *Den, unsigned AtLeast,
1195                                             bool IsSigned) const {
1196   const DataLayout &DL = Mod->getDataLayout();
1197   unsigned LHSSignBits = ComputeNumSignBits(Num, DL, 0, AC, &I);
1198   if (LHSSignBits < AtLeast)
1199     return -1;
1200 
1201   unsigned RHSSignBits = ComputeNumSignBits(Den, DL, 0, AC, &I);
1202   if (RHSSignBits < AtLeast)
1203     return -1;
1204 
1205   unsigned SignBits = std::min(LHSSignBits, RHSSignBits);
1206   unsigned DivBits = Num->getType()->getScalarSizeInBits() - SignBits;
1207   if (IsSigned)
1208     ++DivBits;
1209   return DivBits;
1210 }
1211 
1212 // The fractional part of a float is enough to accurately represent up to
1213 // a 24-bit signed integer.
expandDivRem24(IRBuilder<> & Builder,BinaryOperator & I,Value * Num,Value * Den,bool IsDiv,bool IsSigned) const1214 Value *AMDGPUCodeGenPrepareImpl::expandDivRem24(IRBuilder<> &Builder,
1215                                                 BinaryOperator &I, Value *Num,
1216                                                 Value *Den, bool IsDiv,
1217                                                 bool IsSigned) const {
1218   unsigned SSBits = Num->getType()->getScalarSizeInBits();
1219   // If Num bits <= 24, assume 0 signbits.
1220   unsigned AtLeast = (SSBits <= 24) ? 0 : (SSBits - 24 + IsSigned);
1221   int DivBits = getDivNumBits(I, Num, Den, AtLeast, IsSigned);
1222   if (DivBits == -1)
1223     return nullptr;
1224   return expandDivRem24Impl(Builder, I, Num, Den, DivBits, IsDiv, IsSigned);
1225 }
1226 
expandDivRem24Impl(IRBuilder<> & Builder,BinaryOperator & I,Value * Num,Value * Den,unsigned DivBits,bool IsDiv,bool IsSigned) const1227 Value *AMDGPUCodeGenPrepareImpl::expandDivRem24Impl(
1228     IRBuilder<> &Builder, BinaryOperator &I, Value *Num, Value *Den,
1229     unsigned DivBits, bool IsDiv, bool IsSigned) const {
1230   Type *I32Ty = Builder.getInt32Ty();
1231   Num = Builder.CreateTrunc(Num, I32Ty);
1232   Den = Builder.CreateTrunc(Den, I32Ty);
1233 
1234   Type *F32Ty = Builder.getFloatTy();
1235   ConstantInt *One = Builder.getInt32(1);
1236   Value *JQ = One;
1237 
1238   if (IsSigned) {
1239     // char|short jq = ia ^ ib;
1240     JQ = Builder.CreateXor(Num, Den);
1241 
1242     // jq = jq >> (bitsize - 2)
1243     JQ = Builder.CreateAShr(JQ, Builder.getInt32(30));
1244 
1245     // jq = jq | 0x1
1246     JQ = Builder.CreateOr(JQ, One);
1247   }
1248 
1249   // int ia = (int)LHS;
1250   Value *IA = Num;
1251 
1252   // int ib, (int)RHS;
1253   Value *IB = Den;
1254 
1255   // float fa = (float)ia;
1256   Value *FA = IsSigned ? Builder.CreateSIToFP(IA, F32Ty)
1257                        : Builder.CreateUIToFP(IA, F32Ty);
1258 
1259   // float fb = (float)ib;
1260   Value *FB = IsSigned ? Builder.CreateSIToFP(IB,F32Ty)
1261                        : Builder.CreateUIToFP(IB,F32Ty);
1262 
1263   Function *RcpDecl = Intrinsic::getDeclaration(Mod, Intrinsic::amdgcn_rcp,
1264                                                 Builder.getFloatTy());
1265   Value *RCP = Builder.CreateCall(RcpDecl, { FB });
1266   Value *FQM = Builder.CreateFMul(FA, RCP);
1267 
1268   // fq = trunc(fqm);
1269   CallInst *FQ = Builder.CreateUnaryIntrinsic(Intrinsic::trunc, FQM);
1270   FQ->copyFastMathFlags(Builder.getFastMathFlags());
1271 
1272   // float fqneg = -fq;
1273   Value *FQNeg = Builder.CreateFNeg(FQ);
1274 
1275   // float fr = mad(fqneg, fb, fa);
1276   auto FMAD = !ST->hasMadMacF32Insts()
1277                   ? Intrinsic::fma
1278                   : (Intrinsic::ID)Intrinsic::amdgcn_fmad_ftz;
1279   Value *FR = Builder.CreateIntrinsic(FMAD,
1280                                       {FQNeg->getType()}, {FQNeg, FB, FA}, FQ);
1281 
1282   // int iq = (int)fq;
1283   Value *IQ = IsSigned ? Builder.CreateFPToSI(FQ, I32Ty)
1284                        : Builder.CreateFPToUI(FQ, I32Ty);
1285 
1286   // fr = fabs(fr);
1287   FR = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, FR, FQ);
1288 
1289   // fb = fabs(fb);
1290   FB = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, FB, FQ);
1291 
1292   // int cv = fr >= fb;
1293   Value *CV = Builder.CreateFCmpOGE(FR, FB);
1294 
1295   // jq = (cv ? jq : 0);
1296   JQ = Builder.CreateSelect(CV, JQ, Builder.getInt32(0));
1297 
1298   // dst = iq + jq;
1299   Value *Div = Builder.CreateAdd(IQ, JQ);
1300 
1301   Value *Res = Div;
1302   if (!IsDiv) {
1303     // Rem needs compensation, it's easier to recompute it
1304     Value *Rem = Builder.CreateMul(Div, Den);
1305     Res = Builder.CreateSub(Num, Rem);
1306   }
1307 
1308   if (DivBits != 0 && DivBits < 32) {
1309     // Extend in register from the number of bits this divide really is.
1310     if (IsSigned) {
1311       int InRegBits = 32 - DivBits;
1312 
1313       Res = Builder.CreateShl(Res, InRegBits);
1314       Res = Builder.CreateAShr(Res, InRegBits);
1315     } else {
1316       ConstantInt *TruncMask
1317         = Builder.getInt32((UINT64_C(1) << DivBits) - 1);
1318       Res = Builder.CreateAnd(Res, TruncMask);
1319     }
1320   }
1321 
1322   return Res;
1323 }
1324 
1325 // Try to recognize special cases the DAG will emit special, better expansions
1326 // than the general expansion we do here.
1327 
1328 // TODO: It would be better to just directly handle those optimizations here.
divHasSpecialOptimization(BinaryOperator & I,Value * Num,Value * Den) const1329 bool AMDGPUCodeGenPrepareImpl::divHasSpecialOptimization(BinaryOperator &I,
1330                                                          Value *Num,
1331                                                          Value *Den) const {
1332   if (Constant *C = dyn_cast<Constant>(Den)) {
1333     // Arbitrary constants get a better expansion as long as a wider mulhi is
1334     // legal.
1335     if (C->getType()->getScalarSizeInBits() <= 32)
1336       return true;
1337 
1338     // TODO: Sdiv check for not exact for some reason.
1339 
1340     // If there's no wider mulhi, there's only a better expansion for powers of
1341     // two.
1342     // TODO: Should really know for each vector element.
1343     if (isKnownToBeAPowerOfTwo(C, *DL, true, 0, AC, &I, DT))
1344       return true;
1345 
1346     return false;
1347   }
1348 
1349   if (BinaryOperator *BinOpDen = dyn_cast<BinaryOperator>(Den)) {
1350     // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
1351     if (BinOpDen->getOpcode() == Instruction::Shl &&
1352         isa<Constant>(BinOpDen->getOperand(0)) &&
1353         isKnownToBeAPowerOfTwo(BinOpDen->getOperand(0), *DL, true,
1354                                0, AC, &I, DT)) {
1355       return true;
1356     }
1357   }
1358 
1359   return false;
1360 }
1361 
getSign32(Value * V,IRBuilder<> & Builder,const DataLayout * DL)1362 static Value *getSign32(Value *V, IRBuilder<> &Builder, const DataLayout *DL) {
1363   // Check whether the sign can be determined statically.
1364   KnownBits Known = computeKnownBits(V, *DL);
1365   if (Known.isNegative())
1366     return Constant::getAllOnesValue(V->getType());
1367   if (Known.isNonNegative())
1368     return Constant::getNullValue(V->getType());
1369   return Builder.CreateAShr(V, Builder.getInt32(31));
1370 }
1371 
expandDivRem32(IRBuilder<> & Builder,BinaryOperator & I,Value * X,Value * Y) const1372 Value *AMDGPUCodeGenPrepareImpl::expandDivRem32(IRBuilder<> &Builder,
1373                                                 BinaryOperator &I, Value *X,
1374                                                 Value *Y) const {
1375   Instruction::BinaryOps Opc = I.getOpcode();
1376   assert(Opc == Instruction::URem || Opc == Instruction::UDiv ||
1377          Opc == Instruction::SRem || Opc == Instruction::SDiv);
1378 
1379   FastMathFlags FMF;
1380   FMF.setFast();
1381   Builder.setFastMathFlags(FMF);
1382 
1383   if (divHasSpecialOptimization(I, X, Y))
1384     return nullptr;  // Keep it for later optimization.
1385 
1386   bool IsDiv = Opc == Instruction::UDiv || Opc == Instruction::SDiv;
1387   bool IsSigned = Opc == Instruction::SRem || Opc == Instruction::SDiv;
1388 
1389   Type *Ty = X->getType();
1390   Type *I32Ty = Builder.getInt32Ty();
1391   Type *F32Ty = Builder.getFloatTy();
1392 
1393   if (Ty->getScalarSizeInBits() != 32) {
1394     if (IsSigned) {
1395       X = Builder.CreateSExtOrTrunc(X, I32Ty);
1396       Y = Builder.CreateSExtOrTrunc(Y, I32Ty);
1397     } else {
1398       X = Builder.CreateZExtOrTrunc(X, I32Ty);
1399       Y = Builder.CreateZExtOrTrunc(Y, I32Ty);
1400     }
1401   }
1402 
1403   if (Value *Res = expandDivRem24(Builder, I, X, Y, IsDiv, IsSigned)) {
1404     return IsSigned ? Builder.CreateSExtOrTrunc(Res, Ty) :
1405                       Builder.CreateZExtOrTrunc(Res, Ty);
1406   }
1407 
1408   ConstantInt *Zero = Builder.getInt32(0);
1409   ConstantInt *One = Builder.getInt32(1);
1410 
1411   Value *Sign = nullptr;
1412   if (IsSigned) {
1413     Value *SignX = getSign32(X, Builder, DL);
1414     Value *SignY = getSign32(Y, Builder, DL);
1415     // Remainder sign is the same as LHS
1416     Sign = IsDiv ? Builder.CreateXor(SignX, SignY) : SignX;
1417 
1418     X = Builder.CreateAdd(X, SignX);
1419     Y = Builder.CreateAdd(Y, SignY);
1420 
1421     X = Builder.CreateXor(X, SignX);
1422     Y = Builder.CreateXor(Y, SignY);
1423   }
1424 
1425   // The algorithm here is based on ideas from "Software Integer Division", Tom
1426   // Rodeheffer, August 2008.
1427   //
1428   // unsigned udiv(unsigned x, unsigned y) {
1429   //   // Initial estimate of inv(y). The constant is less than 2^32 to ensure
1430   //   // that this is a lower bound on inv(y), even if some of the calculations
1431   //   // round up.
1432   //   unsigned z = (unsigned)((4294967296.0 - 512.0) * v_rcp_f32((float)y));
1433   //
1434   //   // One round of UNR (Unsigned integer Newton-Raphson) to improve z.
1435   //   // Empirically this is guaranteed to give a "two-y" lower bound on
1436   //   // inv(y).
1437   //   z += umulh(z, -y * z);
1438   //
1439   //   // Quotient/remainder estimate.
1440   //   unsigned q = umulh(x, z);
1441   //   unsigned r = x - q * y;
1442   //
1443   //   // Two rounds of quotient/remainder refinement.
1444   //   if (r >= y) {
1445   //     ++q;
1446   //     r -= y;
1447   //   }
1448   //   if (r >= y) {
1449   //     ++q;
1450   //     r -= y;
1451   //   }
1452   //
1453   //   return q;
1454   // }
1455 
1456   // Initial estimate of inv(y).
1457   Value *FloatY = Builder.CreateUIToFP(Y, F32Ty);
1458   Function *Rcp = Intrinsic::getDeclaration(Mod, Intrinsic::amdgcn_rcp, F32Ty);
1459   Value *RcpY = Builder.CreateCall(Rcp, {FloatY});
1460   Constant *Scale = ConstantFP::get(F32Ty, llvm::bit_cast<float>(0x4F7FFFFE));
1461   Value *ScaledY = Builder.CreateFMul(RcpY, Scale);
1462   Value *Z = Builder.CreateFPToUI(ScaledY, I32Ty);
1463 
1464   // One round of UNR.
1465   Value *NegY = Builder.CreateSub(Zero, Y);
1466   Value *NegYZ = Builder.CreateMul(NegY, Z);
1467   Z = Builder.CreateAdd(Z, getMulHu(Builder, Z, NegYZ));
1468 
1469   // Quotient/remainder estimate.
1470   Value *Q = getMulHu(Builder, X, Z);
1471   Value *R = Builder.CreateSub(X, Builder.CreateMul(Q, Y));
1472 
1473   // First quotient/remainder refinement.
1474   Value *Cond = Builder.CreateICmpUGE(R, Y);
1475   if (IsDiv)
1476     Q = Builder.CreateSelect(Cond, Builder.CreateAdd(Q, One), Q);
1477   R = Builder.CreateSelect(Cond, Builder.CreateSub(R, Y), R);
1478 
1479   // Second quotient/remainder refinement.
1480   Cond = Builder.CreateICmpUGE(R, Y);
1481   Value *Res;
1482   if (IsDiv)
1483     Res = Builder.CreateSelect(Cond, Builder.CreateAdd(Q, One), Q);
1484   else
1485     Res = Builder.CreateSelect(Cond, Builder.CreateSub(R, Y), R);
1486 
1487   if (IsSigned) {
1488     Res = Builder.CreateXor(Res, Sign);
1489     Res = Builder.CreateSub(Res, Sign);
1490     Res = Builder.CreateSExtOrTrunc(Res, Ty);
1491   } else {
1492     Res = Builder.CreateZExtOrTrunc(Res, Ty);
1493   }
1494   return Res;
1495 }
1496 
shrinkDivRem64(IRBuilder<> & Builder,BinaryOperator & I,Value * Num,Value * Den) const1497 Value *AMDGPUCodeGenPrepareImpl::shrinkDivRem64(IRBuilder<> &Builder,
1498                                                 BinaryOperator &I, Value *Num,
1499                                                 Value *Den) const {
1500   if (!ExpandDiv64InIR && divHasSpecialOptimization(I, Num, Den))
1501     return nullptr;  // Keep it for later optimization.
1502 
1503   Instruction::BinaryOps Opc = I.getOpcode();
1504 
1505   bool IsDiv = Opc == Instruction::SDiv || Opc == Instruction::UDiv;
1506   bool IsSigned = Opc == Instruction::SDiv || Opc == Instruction::SRem;
1507 
1508   int NumDivBits = getDivNumBits(I, Num, Den, 32, IsSigned);
1509   if (NumDivBits == -1)
1510     return nullptr;
1511 
1512   Value *Narrowed = nullptr;
1513   if (NumDivBits <= 24) {
1514     Narrowed = expandDivRem24Impl(Builder, I, Num, Den, NumDivBits,
1515                                   IsDiv, IsSigned);
1516   } else if (NumDivBits <= 32) {
1517     Narrowed = expandDivRem32(Builder, I, Num, Den);
1518   }
1519 
1520   if (Narrowed) {
1521     return IsSigned ? Builder.CreateSExt(Narrowed, Num->getType()) :
1522                       Builder.CreateZExt(Narrowed, Num->getType());
1523   }
1524 
1525   return nullptr;
1526 }
1527 
expandDivRem64(BinaryOperator & I) const1528 void AMDGPUCodeGenPrepareImpl::expandDivRem64(BinaryOperator &I) const {
1529   Instruction::BinaryOps Opc = I.getOpcode();
1530   // Do the general expansion.
1531   if (Opc == Instruction::UDiv || Opc == Instruction::SDiv) {
1532     expandDivisionUpTo64Bits(&I);
1533     return;
1534   }
1535 
1536   if (Opc == Instruction::URem || Opc == Instruction::SRem) {
1537     expandRemainderUpTo64Bits(&I);
1538     return;
1539   }
1540 
1541   llvm_unreachable("not a division");
1542 }
1543 
visitBinaryOperator(BinaryOperator & I)1544 bool AMDGPUCodeGenPrepareImpl::visitBinaryOperator(BinaryOperator &I) {
1545   if (foldBinOpIntoSelect(I))
1546     return true;
1547 
1548   if (ST->has16BitInsts() && needsPromotionToI32(I.getType()) &&
1549       UA->isUniform(&I) && promoteUniformOpToI32(I))
1550     return true;
1551 
1552   if (UseMul24Intrin && replaceMulWithMul24(I))
1553     return true;
1554 
1555   bool Changed = false;
1556   Instruction::BinaryOps Opc = I.getOpcode();
1557   Type *Ty = I.getType();
1558   Value *NewDiv = nullptr;
1559   unsigned ScalarSize = Ty->getScalarSizeInBits();
1560 
1561   SmallVector<BinaryOperator *, 8> Div64ToExpand;
1562 
1563   if ((Opc == Instruction::URem || Opc == Instruction::UDiv ||
1564        Opc == Instruction::SRem || Opc == Instruction::SDiv) &&
1565       ScalarSize <= 64 &&
1566       !DisableIDivExpand) {
1567     Value *Num = I.getOperand(0);
1568     Value *Den = I.getOperand(1);
1569     IRBuilder<> Builder(&I);
1570     Builder.SetCurrentDebugLocation(I.getDebugLoc());
1571 
1572     if (auto *VT = dyn_cast<FixedVectorType>(Ty)) {
1573       NewDiv = PoisonValue::get(VT);
1574 
1575       for (unsigned N = 0, E = VT->getNumElements(); N != E; ++N) {
1576         Value *NumEltN = Builder.CreateExtractElement(Num, N);
1577         Value *DenEltN = Builder.CreateExtractElement(Den, N);
1578 
1579         Value *NewElt;
1580         if (ScalarSize <= 32) {
1581           NewElt = expandDivRem32(Builder, I, NumEltN, DenEltN);
1582           if (!NewElt)
1583             NewElt = Builder.CreateBinOp(Opc, NumEltN, DenEltN);
1584         } else {
1585           // See if this 64-bit division can be shrunk to 32/24-bits before
1586           // producing the general expansion.
1587           NewElt = shrinkDivRem64(Builder, I, NumEltN, DenEltN);
1588           if (!NewElt) {
1589             // The general 64-bit expansion introduces control flow and doesn't
1590             // return the new value. Just insert a scalar copy and defer
1591             // expanding it.
1592             NewElt = Builder.CreateBinOp(Opc, NumEltN, DenEltN);
1593             Div64ToExpand.push_back(cast<BinaryOperator>(NewElt));
1594           }
1595         }
1596 
1597         if (auto *NewEltI = dyn_cast<Instruction>(NewElt))
1598           NewEltI->copyIRFlags(&I);
1599 
1600         NewDiv = Builder.CreateInsertElement(NewDiv, NewElt, N);
1601       }
1602     } else {
1603       if (ScalarSize <= 32)
1604         NewDiv = expandDivRem32(Builder, I, Num, Den);
1605       else {
1606         NewDiv = shrinkDivRem64(Builder, I, Num, Den);
1607         if (!NewDiv)
1608           Div64ToExpand.push_back(&I);
1609       }
1610     }
1611 
1612     if (NewDiv) {
1613       I.replaceAllUsesWith(NewDiv);
1614       I.eraseFromParent();
1615       Changed = true;
1616     }
1617   }
1618 
1619   if (ExpandDiv64InIR) {
1620     // TODO: We get much worse code in specially handled constant cases.
1621     for (BinaryOperator *Div : Div64ToExpand) {
1622       expandDivRem64(*Div);
1623       FlowChanged = true;
1624       Changed = true;
1625     }
1626   }
1627 
1628   return Changed;
1629 }
1630 
visitLoadInst(LoadInst & I)1631 bool AMDGPUCodeGenPrepareImpl::visitLoadInst(LoadInst &I) {
1632   if (!WidenLoads)
1633     return false;
1634 
1635   if ((I.getPointerAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS ||
1636        I.getPointerAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS_32BIT) &&
1637       canWidenScalarExtLoad(I)) {
1638     IRBuilder<> Builder(&I);
1639     Builder.SetCurrentDebugLocation(I.getDebugLoc());
1640 
1641     Type *I32Ty = Builder.getInt32Ty();
1642     LoadInst *WidenLoad = Builder.CreateLoad(I32Ty, I.getPointerOperand());
1643     WidenLoad->copyMetadata(I);
1644 
1645     // If we have range metadata, we need to convert the type, and not make
1646     // assumptions about the high bits.
1647     if (auto *Range = WidenLoad->getMetadata(LLVMContext::MD_range)) {
1648       ConstantInt *Lower =
1649         mdconst::extract<ConstantInt>(Range->getOperand(0));
1650 
1651       if (Lower->isNullValue()) {
1652         WidenLoad->setMetadata(LLVMContext::MD_range, nullptr);
1653       } else {
1654         Metadata *LowAndHigh[] = {
1655           ConstantAsMetadata::get(ConstantInt::get(I32Ty, Lower->getValue().zext(32))),
1656           // Don't make assumptions about the high bits.
1657           ConstantAsMetadata::get(ConstantInt::get(I32Ty, 0))
1658         };
1659 
1660         WidenLoad->setMetadata(LLVMContext::MD_range,
1661                                MDNode::get(Mod->getContext(), LowAndHigh));
1662       }
1663     }
1664 
1665     int TySize = Mod->getDataLayout().getTypeSizeInBits(I.getType());
1666     Type *IntNTy = Builder.getIntNTy(TySize);
1667     Value *ValTrunc = Builder.CreateTrunc(WidenLoad, IntNTy);
1668     Value *ValOrig = Builder.CreateBitCast(ValTrunc, I.getType());
1669     I.replaceAllUsesWith(ValOrig);
1670     I.eraseFromParent();
1671     return true;
1672   }
1673 
1674   return false;
1675 }
1676 
visitICmpInst(ICmpInst & I)1677 bool AMDGPUCodeGenPrepareImpl::visitICmpInst(ICmpInst &I) {
1678   bool Changed = false;
1679 
1680   if (ST->has16BitInsts() && needsPromotionToI32(I.getOperand(0)->getType()) &&
1681       UA->isUniform(&I))
1682     Changed |= promoteUniformOpToI32(I);
1683 
1684   return Changed;
1685 }
1686 
visitSelectInst(SelectInst & I)1687 bool AMDGPUCodeGenPrepareImpl::visitSelectInst(SelectInst &I) {
1688   Value *Cond = I.getCondition();
1689   Value *TrueVal = I.getTrueValue();
1690   Value *FalseVal = I.getFalseValue();
1691   Value *CmpVal;
1692   FCmpInst::Predicate Pred;
1693 
1694   if (ST->has16BitInsts() && needsPromotionToI32(I.getType())) {
1695     if (UA->isUniform(&I))
1696       return promoteUniformOpToI32(I);
1697     return false;
1698   }
1699 
1700   // Match fract pattern with nan check.
1701   if (!match(Cond, m_FCmp(Pred, m_Value(CmpVal), m_NonNaN())))
1702     return false;
1703 
1704   FPMathOperator *FPOp = dyn_cast<FPMathOperator>(&I);
1705   if (!FPOp)
1706     return false;
1707 
1708   IRBuilder<> Builder(&I);
1709   Builder.setFastMathFlags(FPOp->getFastMathFlags());
1710 
1711   auto *IITrue = dyn_cast<IntrinsicInst>(TrueVal);
1712   auto *IIFalse = dyn_cast<IntrinsicInst>(FalseVal);
1713 
1714   Value *Fract = nullptr;
1715   if (Pred == FCmpInst::FCMP_UNO && TrueVal == CmpVal && IIFalse &&
1716       CmpVal == matchFractPat(*IIFalse)) {
1717     // isnan(x) ? x : fract(x)
1718     Fract = applyFractPat(Builder, CmpVal);
1719   } else if (Pred == FCmpInst::FCMP_ORD && FalseVal == CmpVal && IITrue &&
1720              CmpVal == matchFractPat(*IITrue)) {
1721     // !isnan(x) ? fract(x) : x
1722     Fract = applyFractPat(Builder, CmpVal);
1723   } else
1724     return false;
1725 
1726   Fract->takeName(&I);
1727   I.replaceAllUsesWith(Fract);
1728   RecursivelyDeleteTriviallyDeadInstructions(&I, TLInfo);
1729   return true;
1730 }
1731 
areInSameBB(const Value * A,const Value * B)1732 static bool areInSameBB(const Value *A, const Value *B) {
1733   const auto *IA = dyn_cast<Instruction>(A);
1734   const auto *IB = dyn_cast<Instruction>(B);
1735   return IA && IB && IA->getParent() == IB->getParent();
1736 }
1737 
1738 // Helper for breaking large PHIs that returns true when an extractelement on V
1739 // is likely to be folded away by the DAG combiner.
isInterestingPHIIncomingValue(const Value * V)1740 static bool isInterestingPHIIncomingValue(const Value *V) {
1741   const auto *FVT = dyn_cast<FixedVectorType>(V->getType());
1742   if (!FVT)
1743     return false;
1744 
1745   const Value *CurVal = V;
1746 
1747   // Check for insertelements, keeping track of the elements covered.
1748   BitVector EltsCovered(FVT->getNumElements());
1749   while (const auto *IE = dyn_cast<InsertElementInst>(CurVal)) {
1750     const auto *Idx = dyn_cast<ConstantInt>(IE->getOperand(2));
1751 
1752     // Non constant index/out of bounds index -> folding is unlikely.
1753     // The latter is more of a sanity check because canonical IR should just
1754     // have replaced those with poison.
1755     if (!Idx || Idx->getZExtValue() >= FVT->getNumElements())
1756       return false;
1757 
1758     const auto *VecSrc = IE->getOperand(0);
1759 
1760     // If the vector source is another instruction, it must be in the same basic
1761     // block. Otherwise, the DAGCombiner won't see the whole thing and is
1762     // unlikely to be able to do anything interesting here.
1763     if (isa<Instruction>(VecSrc) && !areInSameBB(VecSrc, IE))
1764       return false;
1765 
1766     CurVal = VecSrc;
1767     EltsCovered.set(Idx->getZExtValue());
1768 
1769     // All elements covered.
1770     if (EltsCovered.all())
1771       return true;
1772   }
1773 
1774   // We either didn't find a single insertelement, or the insertelement chain
1775   // ended before all elements were covered. Check for other interesting values.
1776 
1777   // Constants are always interesting because we can just constant fold the
1778   // extractelements.
1779   if (isa<Constant>(CurVal))
1780     return true;
1781 
1782   // shufflevector is likely to be profitable if either operand is a constant,
1783   // or if either source is in the same block.
1784   // This is because shufflevector is most often lowered as a series of
1785   // insert/extract elements anyway.
1786   if (const auto *SV = dyn_cast<ShuffleVectorInst>(CurVal)) {
1787     return isa<Constant>(SV->getOperand(1)) ||
1788            areInSameBB(SV, SV->getOperand(0)) ||
1789            areInSameBB(SV, SV->getOperand(1));
1790   }
1791 
1792   return false;
1793 }
1794 
collectPHINodes(const PHINode & I,SmallPtrSet<const PHINode *,8> & SeenPHIs)1795 static void collectPHINodes(const PHINode &I,
1796                             SmallPtrSet<const PHINode *, 8> &SeenPHIs) {
1797   const auto [It, Inserted] = SeenPHIs.insert(&I);
1798   if (!Inserted)
1799     return;
1800 
1801   for (const Value *Inc : I.incoming_values()) {
1802     if (const auto *PhiInc = dyn_cast<PHINode>(Inc))
1803       collectPHINodes(*PhiInc, SeenPHIs);
1804   }
1805 
1806   for (const User *U : I.users()) {
1807     if (const auto *PhiU = dyn_cast<PHINode>(U))
1808       collectPHINodes(*PhiU, SeenPHIs);
1809   }
1810 }
1811 
canBreakPHINode(const PHINode & I)1812 bool AMDGPUCodeGenPrepareImpl::canBreakPHINode(const PHINode &I) {
1813   // Check in the cache first.
1814   if (const auto It = BreakPhiNodesCache.find(&I);
1815       It != BreakPhiNodesCache.end())
1816     return It->second;
1817 
1818   // We consider PHI nodes as part of "chains", so given a PHI node I, we
1819   // recursively consider all its users and incoming values that are also PHI
1820   // nodes. We then make a decision about all of those PHIs at once. Either they
1821   // all get broken up, or none of them do. That way, we avoid cases where a
1822   // single PHI is/is not broken and we end up reforming/exploding a vector
1823   // multiple times, or even worse, doing it in a loop.
1824   SmallPtrSet<const PHINode *, 8> WorkList;
1825   collectPHINodes(I, WorkList);
1826 
1827 #ifndef NDEBUG
1828   // Check that none of the PHI nodes in the worklist are in the map. If some of
1829   // them are, it means we're not good enough at collecting related PHIs.
1830   for (const PHINode *WLP : WorkList) {
1831     assert(BreakPhiNodesCache.count(WLP) == 0);
1832   }
1833 #endif
1834 
1835   // To consider a PHI profitable to break, we need to see some interesting
1836   // incoming values. At least 2/3rd (rounded up) of all PHIs in the worklist
1837   // must have one to consider all PHIs breakable.
1838   //
1839   // This threshold has been determined through performance testing.
1840   //
1841   // Note that the computation below is equivalent to
1842   //
1843   //    (unsigned)ceil((K / 3.0) * 2)
1844   //
1845   // It's simply written this way to avoid mixing integral/FP arithmetic.
1846   const auto Threshold = (alignTo(WorkList.size() * 2, 3) / 3);
1847   unsigned NumBreakablePHIs = 0;
1848   bool CanBreak = false;
1849   for (const PHINode *Cur : WorkList) {
1850     // Don't break PHIs that have no interesting incoming values. That is, where
1851     // there is no clear opportunity to fold the "extractelement" instructions
1852     // we would add.
1853     //
1854     // Note: IC does not run after this pass, so we're only interested in the
1855     // foldings that the DAG combiner can do.
1856     if (any_of(Cur->incoming_values(), isInterestingPHIIncomingValue)) {
1857       if (++NumBreakablePHIs >= Threshold) {
1858         CanBreak = true;
1859         break;
1860       }
1861     }
1862   }
1863 
1864   for (const PHINode *Cur : WorkList)
1865     BreakPhiNodesCache[Cur] = CanBreak;
1866 
1867   return CanBreak;
1868 }
1869 
1870 /// Helper class for "break large PHIs" (visitPHINode).
1871 ///
1872 /// This represents a slice of a PHI's incoming value, which is made up of:
1873 ///   - The type of the slice (Ty)
1874 ///   - The index in the incoming value's vector where the slice starts (Idx)
1875 ///   - The number of elements in the slice (NumElts).
1876 /// It also keeps track of the NewPHI node inserted for this particular slice.
1877 ///
1878 /// Slice examples:
1879 ///   <4 x i64> -> Split into four i64 slices.
1880 ///     -> [i64, 0, 1], [i64, 1, 1], [i64, 2, 1], [i64, 3, 1]
1881 ///   <5 x i16> -> Split into 2 <2 x i16> slices + a i16 tail.
1882 ///     -> [<2 x i16>, 0, 2], [<2 x i16>, 2, 2], [i16, 4, 1]
1883 class VectorSlice {
1884 public:
VectorSlice(Type * Ty,unsigned Idx,unsigned NumElts)1885   VectorSlice(Type *Ty, unsigned Idx, unsigned NumElts)
1886       : Ty(Ty), Idx(Idx), NumElts(NumElts) {}
1887 
1888   Type *Ty = nullptr;
1889   unsigned Idx = 0;
1890   unsigned NumElts = 0;
1891   PHINode *NewPHI = nullptr;
1892 
1893   /// Slice \p Inc according to the information contained within this slice.
1894   /// This is cached, so if called multiple times for the same \p BB & \p Inc
1895   /// pair, it returns the same Sliced value as well.
1896   ///
1897   /// Note this *intentionally* does not return the same value for, say,
1898   /// [%bb.0, %0] & [%bb.1, %0] as:
1899   ///   - It could cause issues with dominance (e.g. if bb.1 is seen first, then
1900   ///   the value in bb.1 may not be reachable from bb.0 if it's its
1901   ///   predecessor.)
1902   ///   - We also want to make our extract instructions as local as possible so
1903   ///   the DAG has better chances of folding them out. Duplicating them like
1904   ///   that is beneficial in that regard.
1905   ///
1906   /// This is both a minor optimization to avoid creating duplicate
1907   /// instructions, but also a requirement for correctness. It is not forbidden
1908   /// for a PHI node to have the same [BB, Val] pair multiple times. If we
1909   /// returned a new value each time, those previously identical pairs would all
1910   /// have different incoming values (from the same block) and it'd cause a "PHI
1911   /// node has multiple entries for the same basic block with different incoming
1912   /// values!" verifier error.
getSlicedVal(BasicBlock * BB,Value * Inc,StringRef NewValName)1913   Value *getSlicedVal(BasicBlock *BB, Value *Inc, StringRef NewValName) {
1914     Value *&Res = SlicedVals[{BB, Inc}];
1915     if (Res)
1916       return Res;
1917 
1918     IRBuilder<> B(BB->getTerminator());
1919     if (Instruction *IncInst = dyn_cast<Instruction>(Inc))
1920       B.SetCurrentDebugLocation(IncInst->getDebugLoc());
1921 
1922     if (NumElts > 1) {
1923       SmallVector<int, 4> Mask;
1924       for (unsigned K = Idx; K < (Idx + NumElts); ++K)
1925         Mask.push_back(K);
1926       Res = B.CreateShuffleVector(Inc, Mask, NewValName);
1927     } else
1928       Res = B.CreateExtractElement(Inc, Idx, NewValName);
1929 
1930     return Res;
1931   }
1932 
1933 private:
1934   SmallDenseMap<std::pair<BasicBlock *, Value *>, Value *> SlicedVals;
1935 };
1936 
visitPHINode(PHINode & I)1937 bool AMDGPUCodeGenPrepareImpl::visitPHINode(PHINode &I) {
1938   // Break-up fixed-vector PHIs into smaller pieces.
1939   // Default threshold is 32, so it breaks up any vector that's >32 bits into
1940   // its elements, or into 32-bit pieces (for 8/16 bit elts).
1941   //
1942   // This is only helpful for DAGISel because it doesn't handle large PHIs as
1943   // well as GlobalISel. DAGISel lowers PHIs by using CopyToReg/CopyFromReg.
1944   // With large, odd-sized PHIs we may end up needing many `build_vector`
1945   // operations with most elements being "undef". This inhibits a lot of
1946   // optimization opportunities and can result in unreasonably high register
1947   // pressure and the inevitable stack spilling.
1948   if (!BreakLargePHIs || getCGPassBuilderOption().EnableGlobalISelOption)
1949     return false;
1950 
1951   FixedVectorType *FVT = dyn_cast<FixedVectorType>(I.getType());
1952   if (!FVT || FVT->getNumElements() == 1 ||
1953       DL->getTypeSizeInBits(FVT) <= BreakLargePHIsThreshold)
1954     return false;
1955 
1956   if (!ForceBreakLargePHIs && !canBreakPHINode(I))
1957     return false;
1958 
1959   std::vector<VectorSlice> Slices;
1960 
1961   Type *EltTy = FVT->getElementType();
1962   {
1963     unsigned Idx = 0;
1964     // For 8/16 bits type, don't scalarize fully but break it up into as many
1965     // 32-bit slices as we can, and scalarize the tail.
1966     const unsigned EltSize = DL->getTypeSizeInBits(EltTy);
1967     const unsigned NumElts = FVT->getNumElements();
1968     if (EltSize == 8 || EltSize == 16) {
1969       const unsigned SubVecSize = (32 / EltSize);
1970       Type *SubVecTy = FixedVectorType::get(EltTy, SubVecSize);
1971       for (unsigned End = alignDown(NumElts, SubVecSize); Idx < End;
1972            Idx += SubVecSize)
1973         Slices.emplace_back(SubVecTy, Idx, SubVecSize);
1974     }
1975 
1976     // Scalarize all remaining elements.
1977     for (; Idx < NumElts; ++Idx)
1978       Slices.emplace_back(EltTy, Idx, 1);
1979   }
1980 
1981   assert(Slices.size() > 1);
1982 
1983   // Create one PHI per vector piece. The "VectorSlice" class takes care of
1984   // creating the necessary instruction to extract the relevant slices of each
1985   // incoming value.
1986   IRBuilder<> B(I.getParent());
1987   B.SetCurrentDebugLocation(I.getDebugLoc());
1988 
1989   unsigned IncNameSuffix = 0;
1990   for (VectorSlice &S : Slices) {
1991     // We need to reset the build on each iteration, because getSlicedVal may
1992     // have inserted something into I's BB.
1993     B.SetInsertPoint(I.getParent()->getFirstNonPHIIt());
1994     S.NewPHI = B.CreatePHI(S.Ty, I.getNumIncomingValues());
1995 
1996     for (const auto &[Idx, BB] : enumerate(I.blocks())) {
1997       S.NewPHI->addIncoming(S.getSlicedVal(BB, I.getIncomingValue(Idx),
1998                                            "largephi.extractslice" +
1999                                                std::to_string(IncNameSuffix++)),
2000                             BB);
2001     }
2002   }
2003 
2004   // And replace this PHI with a vector of all the previous PHI values.
2005   Value *Vec = PoisonValue::get(FVT);
2006   unsigned NameSuffix = 0;
2007   for (VectorSlice &S : Slices) {
2008     const auto ValName = "largephi.insertslice" + std::to_string(NameSuffix++);
2009     if (S.NumElts > 1)
2010       Vec =
2011           B.CreateInsertVector(FVT, Vec, S.NewPHI, B.getInt64(S.Idx), ValName);
2012     else
2013       Vec = B.CreateInsertElement(Vec, S.NewPHI, S.Idx, ValName);
2014   }
2015 
2016   I.replaceAllUsesWith(Vec);
2017   I.eraseFromParent();
2018   return true;
2019 }
2020 
2021 /// \param V  Value to check
2022 /// \param DL DataLayout
2023 /// \param TM TargetMachine (TODO: remove once DL contains nullptr values)
2024 /// \param AS Target Address Space
2025 /// \return true if \p V cannot be the null value of \p AS, false otherwise.
isPtrKnownNeverNull(const Value * V,const DataLayout & DL,const AMDGPUTargetMachine & TM,unsigned AS)2026 static bool isPtrKnownNeverNull(const Value *V, const DataLayout &DL,
2027                                 const AMDGPUTargetMachine &TM, unsigned AS) {
2028   // Pointer cannot be null if it's a block address, GV or alloca.
2029   // NOTE: We don't support extern_weak, but if we did, we'd need to check for
2030   // it as the symbol could be null in such cases.
2031   if (isa<BlockAddress>(V) || isa<GlobalValue>(V) || isa<AllocaInst>(V))
2032     return true;
2033 
2034   // Check nonnull arguments.
2035   if (const auto *Arg = dyn_cast<Argument>(V); Arg && Arg->hasNonNullAttr())
2036     return true;
2037 
2038   // getUnderlyingObject may have looked through another addrspacecast, although
2039   // the optimizable situations most likely folded out by now.
2040   if (AS != cast<PointerType>(V->getType())->getAddressSpace())
2041     return false;
2042 
2043   // TODO: Calls that return nonnull?
2044 
2045   // For all other things, use KnownBits.
2046   // We either use 0 or all bits set to indicate null, so check whether the
2047   // value can be zero or all ones.
2048   //
2049   // TODO: Use ValueTracking's isKnownNeverNull if it becomes aware that some
2050   // address spaces have non-zero null values.
2051   auto SrcPtrKB = computeKnownBits(V, DL);
2052   const auto NullVal = TM.getNullPointerValue(AS);
2053 
2054   assert(SrcPtrKB.getBitWidth() == DL.getPointerSizeInBits(AS));
2055   assert((NullVal == 0 || NullVal == -1) &&
2056          "don't know how to check for this null value!");
2057   return NullVal ? !SrcPtrKB.getMaxValue().isAllOnes() : SrcPtrKB.isNonZero();
2058 }
2059 
visitAddrSpaceCastInst(AddrSpaceCastInst & I)2060 bool AMDGPUCodeGenPrepareImpl::visitAddrSpaceCastInst(AddrSpaceCastInst &I) {
2061   // Intrinsic doesn't support vectors, also it seems that it's often difficult
2062   // to prove that a vector cannot have any nulls in it so it's unclear if it's
2063   // worth supporting.
2064   if (I.getType()->isVectorTy())
2065     return false;
2066 
2067   // Check if this can be lowered to a amdgcn.addrspacecast.nonnull.
2068   // This is only worthwhile for casts from/to priv/local to flat.
2069   const unsigned SrcAS = I.getSrcAddressSpace();
2070   const unsigned DstAS = I.getDestAddressSpace();
2071 
2072   bool CanLower = false;
2073   if (SrcAS == AMDGPUAS::FLAT_ADDRESS)
2074     CanLower = (DstAS == AMDGPUAS::LOCAL_ADDRESS ||
2075                 DstAS == AMDGPUAS::PRIVATE_ADDRESS);
2076   else if (DstAS == AMDGPUAS::FLAT_ADDRESS)
2077     CanLower = (SrcAS == AMDGPUAS::LOCAL_ADDRESS ||
2078                 SrcAS == AMDGPUAS::PRIVATE_ADDRESS);
2079   if (!CanLower)
2080     return false;
2081 
2082   SmallVector<const Value *, 4> WorkList;
2083   getUnderlyingObjects(I.getOperand(0), WorkList);
2084   if (!all_of(WorkList, [&](const Value *V) {
2085         return isPtrKnownNeverNull(V, *DL, *TM, SrcAS);
2086       }))
2087     return false;
2088 
2089   IRBuilder<> B(&I);
2090   auto *Intrin = B.CreateIntrinsic(
2091       I.getType(), Intrinsic::amdgcn_addrspacecast_nonnull, {I.getOperand(0)});
2092   I.replaceAllUsesWith(Intrin);
2093   I.eraseFromParent();
2094   return true;
2095 }
2096 
visitIntrinsicInst(IntrinsicInst & I)2097 bool AMDGPUCodeGenPrepareImpl::visitIntrinsicInst(IntrinsicInst &I) {
2098   switch (I.getIntrinsicID()) {
2099   case Intrinsic::bitreverse:
2100     return visitBitreverseIntrinsicInst(I);
2101   case Intrinsic::minnum:
2102     return visitMinNum(I);
2103   case Intrinsic::sqrt:
2104     return visitSqrt(I);
2105   default:
2106     return false;
2107   }
2108 }
2109 
visitBitreverseIntrinsicInst(IntrinsicInst & I)2110 bool AMDGPUCodeGenPrepareImpl::visitBitreverseIntrinsicInst(IntrinsicInst &I) {
2111   bool Changed = false;
2112 
2113   if (ST->has16BitInsts() && needsPromotionToI32(I.getType()) &&
2114       UA->isUniform(&I))
2115     Changed |= promoteUniformBitreverseToI32(I);
2116 
2117   return Changed;
2118 }
2119 
2120 /// Match non-nan fract pattern.
2121 ///   minnum(fsub(x, floor(x)), nextafter(1.0, -1.0)
2122 ///
2123 /// If fract is a useful instruction for the subtarget. Does not account for the
2124 /// nan handling; the instruction has a nan check on the input value.
matchFractPat(IntrinsicInst & I)2125 Value *AMDGPUCodeGenPrepareImpl::matchFractPat(IntrinsicInst &I) {
2126   if (ST->hasFractBug())
2127     return nullptr;
2128 
2129   if (I.getIntrinsicID() != Intrinsic::minnum)
2130     return nullptr;
2131 
2132   Type *Ty = I.getType();
2133   if (!isLegalFloatingTy(Ty->getScalarType()))
2134     return nullptr;
2135 
2136   Value *Arg0 = I.getArgOperand(0);
2137   Value *Arg1 = I.getArgOperand(1);
2138 
2139   const APFloat *C;
2140   if (!match(Arg1, m_APFloat(C)))
2141     return nullptr;
2142 
2143   APFloat One(1.0);
2144   bool LosesInfo;
2145   One.convert(C->getSemantics(), APFloat::rmNearestTiesToEven, &LosesInfo);
2146 
2147   // Match nextafter(1.0, -1)
2148   One.next(true);
2149   if (One != *C)
2150     return nullptr;
2151 
2152   Value *FloorSrc;
2153   if (match(Arg0, m_FSub(m_Value(FloorSrc),
2154                          m_Intrinsic<Intrinsic::floor>(m_Deferred(FloorSrc)))))
2155     return FloorSrc;
2156   return nullptr;
2157 }
2158 
applyFractPat(IRBuilder<> & Builder,Value * FractArg)2159 Value *AMDGPUCodeGenPrepareImpl::applyFractPat(IRBuilder<> &Builder,
2160                                                Value *FractArg) {
2161   SmallVector<Value *, 4> FractVals;
2162   extractValues(Builder, FractVals, FractArg);
2163 
2164   SmallVector<Value *, 4> ResultVals(FractVals.size());
2165 
2166   Type *Ty = FractArg->getType()->getScalarType();
2167   for (unsigned I = 0, E = FractVals.size(); I != E; ++I) {
2168     ResultVals[I] =
2169         Builder.CreateIntrinsic(Intrinsic::amdgcn_fract, {Ty}, {FractVals[I]});
2170   }
2171 
2172   return insertValues(Builder, FractArg->getType(), ResultVals);
2173 }
2174 
visitMinNum(IntrinsicInst & I)2175 bool AMDGPUCodeGenPrepareImpl::visitMinNum(IntrinsicInst &I) {
2176   Value *FractArg = matchFractPat(I);
2177   if (!FractArg)
2178     return false;
2179 
2180   // Match pattern for fract intrinsic in contexts where the nan check has been
2181   // optimized out (and hope the knowledge the source can't be nan wasn't lost).
2182   if (!I.hasNoNaNs() &&
2183       !isKnownNeverNaN(FractArg, /*Depth=*/0, SimplifyQuery(*DL, TLInfo)))
2184     return false;
2185 
2186   IRBuilder<> Builder(&I);
2187   FastMathFlags FMF = I.getFastMathFlags();
2188   FMF.setNoNaNs();
2189   Builder.setFastMathFlags(FMF);
2190 
2191   Value *Fract = applyFractPat(Builder, FractArg);
2192   Fract->takeName(&I);
2193   I.replaceAllUsesWith(Fract);
2194 
2195   RecursivelyDeleteTriviallyDeadInstructions(&I, TLInfo);
2196   return true;
2197 }
2198 
isOneOrNegOne(const Value * Val)2199 static bool isOneOrNegOne(const Value *Val) {
2200   const APFloat *C;
2201   return match(Val, m_APFloat(C)) && C->getExactLog2Abs() == 0;
2202 }
2203 
2204 // Expand llvm.sqrt.f32 calls with !fpmath metadata in a semi-fast way.
visitSqrt(IntrinsicInst & Sqrt)2205 bool AMDGPUCodeGenPrepareImpl::visitSqrt(IntrinsicInst &Sqrt) {
2206   Type *Ty = Sqrt.getType()->getScalarType();
2207   if (!Ty->isFloatTy() && (!Ty->isHalfTy() || ST->has16BitInsts()))
2208     return false;
2209 
2210   const FPMathOperator *FPOp = cast<const FPMathOperator>(&Sqrt);
2211   FastMathFlags SqrtFMF = FPOp->getFastMathFlags();
2212 
2213   // We're trying to handle the fast-but-not-that-fast case only. The lowering
2214   // of fast llvm.sqrt will give the raw instruction anyway.
2215   if (SqrtFMF.approxFunc() || HasUnsafeFPMath)
2216     return false;
2217 
2218   const float ReqdAccuracy = FPOp->getFPAccuracy();
2219 
2220   // Defer correctly rounded expansion to codegen.
2221   if (ReqdAccuracy < 1.0f)
2222     return false;
2223 
2224   // FIXME: This is an ugly hack for this pass using forward iteration instead
2225   // of reverse. If it worked like a normal combiner, the rsq would form before
2226   // we saw a sqrt call.
2227   auto *FDiv =
2228       dyn_cast_or_null<FPMathOperator>(Sqrt.getUniqueUndroppableUser());
2229   if (FDiv && FDiv->getOpcode() == Instruction::FDiv &&
2230       FDiv->getFPAccuracy() >= 1.0f &&
2231       canOptimizeWithRsq(FPOp, FDiv->getFastMathFlags(), SqrtFMF) &&
2232       // TODO: We should also handle the arcp case for the fdiv with non-1 value
2233       isOneOrNegOne(FDiv->getOperand(0)))
2234     return false;
2235 
2236   Value *SrcVal = Sqrt.getOperand(0);
2237   bool CanTreatAsDAZ = canIgnoreDenormalInput(SrcVal, &Sqrt);
2238 
2239   // The raw instruction is 1 ulp, but the correction for denormal handling
2240   // brings it to 2.
2241   if (!CanTreatAsDAZ && ReqdAccuracy < 2.0f)
2242     return false;
2243 
2244   IRBuilder<> Builder(&Sqrt);
2245   SmallVector<Value *, 4> SrcVals;
2246   extractValues(Builder, SrcVals, SrcVal);
2247 
2248   SmallVector<Value *, 4> ResultVals(SrcVals.size());
2249   for (int I = 0, E = SrcVals.size(); I != E; ++I) {
2250     if (CanTreatAsDAZ)
2251       ResultVals[I] = Builder.CreateCall(getSqrtF32(), SrcVals[I]);
2252     else
2253       ResultVals[I] = emitSqrtIEEE2ULP(Builder, SrcVals[I], SqrtFMF);
2254   }
2255 
2256   Value *NewSqrt = insertValues(Builder, Sqrt.getType(), ResultVals);
2257   NewSqrt->takeName(&Sqrt);
2258   Sqrt.replaceAllUsesWith(NewSqrt);
2259   Sqrt.eraseFromParent();
2260   return true;
2261 }
2262 
doInitialization(Module & M)2263 bool AMDGPUCodeGenPrepare::doInitialization(Module &M) {
2264   Impl.Mod = &M;
2265   Impl.DL = &Impl.Mod->getDataLayout();
2266   Impl.SqrtF32 = nullptr;
2267   Impl.LdexpF32 = nullptr;
2268   return false;
2269 }
2270 
runOnFunction(Function & F)2271 bool AMDGPUCodeGenPrepare::runOnFunction(Function &F) {
2272   if (skipFunction(F))
2273     return false;
2274 
2275   auto *TPC = getAnalysisIfAvailable<TargetPassConfig>();
2276   if (!TPC)
2277     return false;
2278 
2279   const AMDGPUTargetMachine &TM = TPC->getTM<AMDGPUTargetMachine>();
2280   Impl.TM = &TM;
2281   Impl.TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
2282   Impl.ST = &TM.getSubtarget<GCNSubtarget>(F);
2283   Impl.AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2284   Impl.UA = &getAnalysis<UniformityInfoWrapperPass>().getUniformityInfo();
2285   auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>();
2286   Impl.DT = DTWP ? &DTWP->getDomTree() : nullptr;
2287   Impl.HasUnsafeFPMath = hasUnsafeFPMath(F);
2288   SIModeRegisterDefaults Mode(F, *Impl.ST);
2289   Impl.HasFP32DenormalFlush =
2290       Mode.FP32Denormals == DenormalMode::getPreserveSign();
2291   return Impl.run(F);
2292 }
2293 
run(Function & F,FunctionAnalysisManager & FAM)2294 PreservedAnalyses AMDGPUCodeGenPreparePass::run(Function &F,
2295                                                 FunctionAnalysisManager &FAM) {
2296   AMDGPUCodeGenPrepareImpl Impl;
2297   Impl.Mod = F.getParent();
2298   Impl.DL = &Impl.Mod->getDataLayout();
2299   Impl.TM = static_cast<const AMDGPUTargetMachine *>(&TM);
2300   Impl.TLInfo = &FAM.getResult<TargetLibraryAnalysis>(F);
2301   Impl.ST = &TM.getSubtarget<GCNSubtarget>(F);
2302   Impl.AC = &FAM.getResult<AssumptionAnalysis>(F);
2303   Impl.UA = &FAM.getResult<UniformityInfoAnalysis>(F);
2304   Impl.DT = FAM.getCachedResult<DominatorTreeAnalysis>(F);
2305   Impl.HasUnsafeFPMath = hasUnsafeFPMath(F);
2306   SIModeRegisterDefaults Mode(F, *Impl.ST);
2307   Impl.HasFP32DenormalFlush =
2308       Mode.FP32Denormals == DenormalMode::getPreserveSign();
2309   PreservedAnalyses PA = PreservedAnalyses::none();
2310   if (!Impl.FlowChanged)
2311     PA.preserveSet<CFGAnalyses>();
2312   return Impl.run(F) ? PA : PreservedAnalyses::all();
2313 }
2314 
2315 INITIALIZE_PASS_BEGIN(AMDGPUCodeGenPrepare, DEBUG_TYPE,
2316                       "AMDGPU IR optimizations", false, false)
2317 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
2318 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
2319 INITIALIZE_PASS_DEPENDENCY(UniformityInfoWrapperPass)
2320 INITIALIZE_PASS_END(AMDGPUCodeGenPrepare, DEBUG_TYPE, "AMDGPU IR optimizations",
2321                     false, false)
2322 
2323 char AMDGPUCodeGenPrepare::ID = 0;
2324 
createAMDGPUCodeGenPreparePass()2325 FunctionPass *llvm::createAMDGPUCodeGenPreparePass() {
2326   return new AMDGPUCodeGenPrepare();
2327 }
2328