xref: /llvm-project/llvm/lib/Target/AMDGPU/AMDGPUAtomicOptimizer.cpp (revision 8e702735090388a3231a863e343f880d0f96fecb)
1 //===-- AMDGPUAtomicOptimizer.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 optimizes atomic operations by using a single lane of a wavefront
11 /// to perform the atomic operation, thus reducing contention on that memory
12 /// location.
13 /// Atomic optimizer uses following strategies to compute scan and reduced
14 /// values
15 /// 1. DPP -
16 ///   This is the most efficient implementation for scan. DPP uses Whole Wave
17 ///   Mode (WWM)
18 /// 2. Iterative -
19 //    An alternative implementation iterates over all active lanes
20 ///   of Wavefront using llvm.cttz and performs scan  using readlane & writelane
21 ///   intrinsics
22 //===----------------------------------------------------------------------===//
23 
24 #include "AMDGPU.h"
25 #include "GCNSubtarget.h"
26 #include "llvm/Analysis/DomTreeUpdater.h"
27 #include "llvm/Analysis/UniformityAnalysis.h"
28 #include "llvm/CodeGen/TargetPassConfig.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/InstVisitor.h"
31 #include "llvm/IR/IntrinsicsAMDGPU.h"
32 #include "llvm/InitializePasses.h"
33 #include "llvm/Target/TargetMachine.h"
34 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
35 
36 #define DEBUG_TYPE "amdgpu-atomic-optimizer"
37 
38 using namespace llvm;
39 using namespace llvm::AMDGPU;
40 
41 namespace {
42 
43 struct ReplacementInfo {
44   Instruction *I;
45   AtomicRMWInst::BinOp Op;
46   unsigned ValIdx;
47   bool ValDivergent;
48 };
49 
50 class AMDGPUAtomicOptimizer : public FunctionPass {
51 public:
52   static char ID;
53   ScanOptions ScanImpl;
54   AMDGPUAtomicOptimizer(ScanOptions ScanImpl)
55       : FunctionPass(ID), ScanImpl(ScanImpl) {}
56 
57   bool runOnFunction(Function &F) override;
58 
59   void getAnalysisUsage(AnalysisUsage &AU) const override {
60     AU.addPreserved<DominatorTreeWrapperPass>();
61     AU.addRequired<UniformityInfoWrapperPass>();
62     AU.addRequired<TargetPassConfig>();
63   }
64 };
65 
66 class AMDGPUAtomicOptimizerImpl
67     : public InstVisitor<AMDGPUAtomicOptimizerImpl> {
68 private:
69   Function &F;
70   SmallVector<ReplacementInfo, 8> ToReplace;
71   const UniformityInfo &UA;
72   const DataLayout &DL;
73   DomTreeUpdater &DTU;
74   const GCNSubtarget &ST;
75   bool IsPixelShader;
76   ScanOptions ScanImpl;
77 
78   Value *buildReduction(IRBuilder<> &B, AtomicRMWInst::BinOp Op, Value *V,
79                         Value *const Identity) const;
80   Value *buildScan(IRBuilder<> &B, AtomicRMWInst::BinOp Op, Value *V,
81                    Value *const Identity) const;
82   Value *buildShiftRight(IRBuilder<> &B, Value *V, Value *const Identity) const;
83 
84   std::pair<Value *, Value *>
85   buildScanIteratively(IRBuilder<> &B, AtomicRMWInst::BinOp Op,
86                        Value *const Identity, Value *V, Instruction &I,
87                        BasicBlock *ComputeLoop, BasicBlock *ComputeEnd) const;
88 
89   void optimizeAtomic(Instruction &I, AtomicRMWInst::BinOp Op, unsigned ValIdx,
90                       bool ValDivergent) const;
91 
92 public:
93   AMDGPUAtomicOptimizerImpl() = delete;
94 
95   AMDGPUAtomicOptimizerImpl(Function &F, const UniformityInfo &UA,
96                             DomTreeUpdater &DTU, const GCNSubtarget &ST,
97                             ScanOptions ScanImpl)
98       : F(F), UA(UA), DL(F.getDataLayout()), DTU(DTU), ST(ST),
99         IsPixelShader(F.getCallingConv() == CallingConv::AMDGPU_PS),
100         ScanImpl(ScanImpl) {}
101 
102   bool run();
103 
104   void visitAtomicRMWInst(AtomicRMWInst &I);
105   void visitIntrinsicInst(IntrinsicInst &I);
106 };
107 
108 } // namespace
109 
110 char AMDGPUAtomicOptimizer::ID = 0;
111 
112 char &llvm::AMDGPUAtomicOptimizerID = AMDGPUAtomicOptimizer::ID;
113 
114 bool AMDGPUAtomicOptimizer::runOnFunction(Function &F) {
115   if (skipFunction(F)) {
116     return false;
117   }
118 
119   const UniformityInfo &UA =
120       getAnalysis<UniformityInfoWrapperPass>().getUniformityInfo();
121 
122   DominatorTreeWrapperPass *DTW =
123       getAnalysisIfAvailable<DominatorTreeWrapperPass>();
124   DomTreeUpdater DTU(DTW ? &DTW->getDomTree() : nullptr,
125                      DomTreeUpdater::UpdateStrategy::Lazy);
126 
127   const TargetPassConfig &TPC = getAnalysis<TargetPassConfig>();
128   const TargetMachine &TM = TPC.getTM<TargetMachine>();
129   const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F);
130 
131   return AMDGPUAtomicOptimizerImpl(F, UA, DTU, ST, ScanImpl).run();
132 }
133 
134 PreservedAnalyses AMDGPUAtomicOptimizerPass::run(Function &F,
135                                                  FunctionAnalysisManager &AM) {
136   const auto &UA = AM.getResult<UniformityInfoAnalysis>(F);
137 
138   DomTreeUpdater DTU(&AM.getResult<DominatorTreeAnalysis>(F),
139                      DomTreeUpdater::UpdateStrategy::Lazy);
140   const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F);
141 
142   bool IsChanged = AMDGPUAtomicOptimizerImpl(F, UA, DTU, ST, ScanImpl).run();
143 
144   if (!IsChanged) {
145     return PreservedAnalyses::all();
146   }
147 
148   PreservedAnalyses PA;
149   PA.preserve<DominatorTreeAnalysis>();
150   return PA;
151 }
152 
153 bool AMDGPUAtomicOptimizerImpl::run() {
154 
155   // Scan option None disables the Pass
156   if (ScanImpl == ScanOptions::None) {
157     return false;
158   }
159 
160   visit(F);
161 
162   const bool Changed = !ToReplace.empty();
163 
164   for (ReplacementInfo &Info : ToReplace) {
165     optimizeAtomic(*Info.I, Info.Op, Info.ValIdx, Info.ValDivergent);
166   }
167 
168   ToReplace.clear();
169 
170   return Changed;
171 }
172 
173 static bool isLegalCrossLaneType(Type *Ty) {
174   switch (Ty->getTypeID()) {
175   case Type::FloatTyID:
176   case Type::DoubleTyID:
177     return true;
178   case Type::IntegerTyID: {
179     unsigned Size = Ty->getIntegerBitWidth();
180     return (Size == 32 || Size == 64);
181   }
182   default:
183     return false;
184   }
185 }
186 
187 void AMDGPUAtomicOptimizerImpl::visitAtomicRMWInst(AtomicRMWInst &I) {
188   // Early exit for unhandled address space atomic instructions.
189   switch (I.getPointerAddressSpace()) {
190   default:
191     return;
192   case AMDGPUAS::GLOBAL_ADDRESS:
193   case AMDGPUAS::LOCAL_ADDRESS:
194     break;
195   }
196 
197   AtomicRMWInst::BinOp Op = I.getOperation();
198 
199   switch (Op) {
200   default:
201     return;
202   case AtomicRMWInst::Add:
203   case AtomicRMWInst::Sub:
204   case AtomicRMWInst::And:
205   case AtomicRMWInst::Or:
206   case AtomicRMWInst::Xor:
207   case AtomicRMWInst::Max:
208   case AtomicRMWInst::Min:
209   case AtomicRMWInst::UMax:
210   case AtomicRMWInst::UMin:
211   case AtomicRMWInst::FAdd:
212   case AtomicRMWInst::FSub:
213   case AtomicRMWInst::FMax:
214   case AtomicRMWInst::FMin:
215     break;
216   }
217 
218   // Only 32 and 64 bit floating point atomic ops are supported.
219   if (AtomicRMWInst::isFPOperation(Op) &&
220       !(I.getType()->isFloatTy() || I.getType()->isDoubleTy())) {
221     return;
222   }
223 
224   const unsigned PtrIdx = 0;
225   const unsigned ValIdx = 1;
226 
227   // If the pointer operand is divergent, then each lane is doing an atomic
228   // operation on a different address, and we cannot optimize that.
229   if (UA.isDivergentUse(I.getOperandUse(PtrIdx))) {
230     return;
231   }
232 
233   bool ValDivergent = UA.isDivergentUse(I.getOperandUse(ValIdx));
234 
235   // If the value operand is divergent, each lane is contributing a different
236   // value to the atomic calculation. We can only optimize divergent values if
237   // we have DPP available on our subtarget (for DPP strategy), and the atomic
238   // operation is 32 or 64 bits.
239   if (ValDivergent) {
240     if (ScanImpl == ScanOptions::DPP && !ST.hasDPP())
241       return;
242 
243     if (!isLegalCrossLaneType(I.getType()))
244       return;
245   }
246 
247   // If we get here, we can optimize the atomic using a single wavefront-wide
248   // atomic operation to do the calculation for the entire wavefront, so
249   // remember the instruction so we can come back to it.
250   const ReplacementInfo Info = {&I, Op, ValIdx, ValDivergent};
251 
252   ToReplace.push_back(Info);
253 }
254 
255 void AMDGPUAtomicOptimizerImpl::visitIntrinsicInst(IntrinsicInst &I) {
256   AtomicRMWInst::BinOp Op;
257 
258   switch (I.getIntrinsicID()) {
259   default:
260     return;
261   case Intrinsic::amdgcn_struct_buffer_atomic_add:
262   case Intrinsic::amdgcn_struct_ptr_buffer_atomic_add:
263   case Intrinsic::amdgcn_raw_buffer_atomic_add:
264   case Intrinsic::amdgcn_raw_ptr_buffer_atomic_add:
265     Op = AtomicRMWInst::Add;
266     break;
267   case Intrinsic::amdgcn_struct_buffer_atomic_sub:
268   case Intrinsic::amdgcn_struct_ptr_buffer_atomic_sub:
269   case Intrinsic::amdgcn_raw_buffer_atomic_sub:
270   case Intrinsic::amdgcn_raw_ptr_buffer_atomic_sub:
271     Op = AtomicRMWInst::Sub;
272     break;
273   case Intrinsic::amdgcn_struct_buffer_atomic_and:
274   case Intrinsic::amdgcn_struct_ptr_buffer_atomic_and:
275   case Intrinsic::amdgcn_raw_buffer_atomic_and:
276   case Intrinsic::amdgcn_raw_ptr_buffer_atomic_and:
277     Op = AtomicRMWInst::And;
278     break;
279   case Intrinsic::amdgcn_struct_buffer_atomic_or:
280   case Intrinsic::amdgcn_struct_ptr_buffer_atomic_or:
281   case Intrinsic::amdgcn_raw_buffer_atomic_or:
282   case Intrinsic::amdgcn_raw_ptr_buffer_atomic_or:
283     Op = AtomicRMWInst::Or;
284     break;
285   case Intrinsic::amdgcn_struct_buffer_atomic_xor:
286   case Intrinsic::amdgcn_struct_ptr_buffer_atomic_xor:
287   case Intrinsic::amdgcn_raw_buffer_atomic_xor:
288   case Intrinsic::amdgcn_raw_ptr_buffer_atomic_xor:
289     Op = AtomicRMWInst::Xor;
290     break;
291   case Intrinsic::amdgcn_struct_buffer_atomic_smin:
292   case Intrinsic::amdgcn_struct_ptr_buffer_atomic_smin:
293   case Intrinsic::amdgcn_raw_buffer_atomic_smin:
294   case Intrinsic::amdgcn_raw_ptr_buffer_atomic_smin:
295     Op = AtomicRMWInst::Min;
296     break;
297   case Intrinsic::amdgcn_struct_buffer_atomic_umin:
298   case Intrinsic::amdgcn_struct_ptr_buffer_atomic_umin:
299   case Intrinsic::amdgcn_raw_buffer_atomic_umin:
300   case Intrinsic::amdgcn_raw_ptr_buffer_atomic_umin:
301     Op = AtomicRMWInst::UMin;
302     break;
303   case Intrinsic::amdgcn_struct_buffer_atomic_smax:
304   case Intrinsic::amdgcn_struct_ptr_buffer_atomic_smax:
305   case Intrinsic::amdgcn_raw_buffer_atomic_smax:
306   case Intrinsic::amdgcn_raw_ptr_buffer_atomic_smax:
307     Op = AtomicRMWInst::Max;
308     break;
309   case Intrinsic::amdgcn_struct_buffer_atomic_umax:
310   case Intrinsic::amdgcn_struct_ptr_buffer_atomic_umax:
311   case Intrinsic::amdgcn_raw_buffer_atomic_umax:
312   case Intrinsic::amdgcn_raw_ptr_buffer_atomic_umax:
313     Op = AtomicRMWInst::UMax;
314     break;
315   }
316 
317   const unsigned ValIdx = 0;
318 
319   const bool ValDivergent = UA.isDivergentUse(I.getOperandUse(ValIdx));
320 
321   // If the value operand is divergent, each lane is contributing a different
322   // value to the atomic calculation. We can only optimize divergent values if
323   // we have DPP available on our subtarget (for DPP strategy), and the atomic
324   // operation is 32 or 64 bits.
325   if (ValDivergent) {
326     if (ScanImpl == ScanOptions::DPP && !ST.hasDPP())
327       return;
328 
329     if (!isLegalCrossLaneType(I.getType()))
330       return;
331   }
332 
333   // If any of the other arguments to the intrinsic are divergent, we can't
334   // optimize the operation.
335   for (unsigned Idx = 1; Idx < I.getNumOperands(); Idx++) {
336     if (UA.isDivergentUse(I.getOperandUse(Idx))) {
337       return;
338     }
339   }
340 
341   // If we get here, we can optimize the atomic using a single wavefront-wide
342   // atomic operation to do the calculation for the entire wavefront, so
343   // remember the instruction so we can come back to it.
344   const ReplacementInfo Info = {&I, Op, ValIdx, ValDivergent};
345 
346   ToReplace.push_back(Info);
347 }
348 
349 // Use the builder to create the non-atomic counterpart of the specified
350 // atomicrmw binary op.
351 static Value *buildNonAtomicBinOp(IRBuilder<> &B, AtomicRMWInst::BinOp Op,
352                                   Value *LHS, Value *RHS) {
353   CmpInst::Predicate Pred;
354 
355   switch (Op) {
356   default:
357     llvm_unreachable("Unhandled atomic op");
358   case AtomicRMWInst::Add:
359     return B.CreateBinOp(Instruction::Add, LHS, RHS);
360   case AtomicRMWInst::FAdd:
361     return B.CreateFAdd(LHS, RHS);
362   case AtomicRMWInst::Sub:
363     return B.CreateBinOp(Instruction::Sub, LHS, RHS);
364   case AtomicRMWInst::FSub:
365     return B.CreateFSub(LHS, RHS);
366   case AtomicRMWInst::And:
367     return B.CreateBinOp(Instruction::And, LHS, RHS);
368   case AtomicRMWInst::Or:
369     return B.CreateBinOp(Instruction::Or, LHS, RHS);
370   case AtomicRMWInst::Xor:
371     return B.CreateBinOp(Instruction::Xor, LHS, RHS);
372 
373   case AtomicRMWInst::Max:
374     Pred = CmpInst::ICMP_SGT;
375     break;
376   case AtomicRMWInst::Min:
377     Pred = CmpInst::ICMP_SLT;
378     break;
379   case AtomicRMWInst::UMax:
380     Pred = CmpInst::ICMP_UGT;
381     break;
382   case AtomicRMWInst::UMin:
383     Pred = CmpInst::ICMP_ULT;
384     break;
385   case AtomicRMWInst::FMax:
386     return B.CreateMaxNum(LHS, RHS);
387   case AtomicRMWInst::FMin:
388     return B.CreateMinNum(LHS, RHS);
389   }
390   Value *Cond = B.CreateICmp(Pred, LHS, RHS);
391   return B.CreateSelect(Cond, LHS, RHS);
392 }
393 
394 // Use the builder to create a reduction of V across the wavefront, with all
395 // lanes active, returning the same result in all lanes.
396 Value *AMDGPUAtomicOptimizerImpl::buildReduction(IRBuilder<> &B,
397                                                  AtomicRMWInst::BinOp Op,
398                                                  Value *V,
399                                                  Value *const Identity) const {
400   Type *AtomicTy = V->getType();
401   Module *M = B.GetInsertBlock()->getModule();
402 
403   // Reduce within each row of 16 lanes.
404   for (unsigned Idx = 0; Idx < 4; Idx++) {
405     V = buildNonAtomicBinOp(
406         B, Op, V,
407         B.CreateIntrinsic(Intrinsic::amdgcn_update_dpp, AtomicTy,
408                           {Identity, V, B.getInt32(DPP::ROW_XMASK0 | 1 << Idx),
409                            B.getInt32(0xf), B.getInt32(0xf), B.getFalse()}));
410   }
411 
412   // Reduce within each pair of rows (i.e. 32 lanes).
413   assert(ST.hasPermLaneX16());
414   Value *Permlanex16Call =
415       B.CreateIntrinsic(AtomicTy, Intrinsic::amdgcn_permlanex16,
416                         {PoisonValue::get(AtomicTy), V, B.getInt32(0),
417                          B.getInt32(0), B.getFalse(), B.getFalse()});
418   V = buildNonAtomicBinOp(B, Op, V, Permlanex16Call);
419   if (ST.isWave32()) {
420     return V;
421   }
422 
423   if (ST.hasPermLane64()) {
424     // Reduce across the upper and lower 32 lanes.
425     Value *Permlane64Call =
426         B.CreateIntrinsic(AtomicTy, Intrinsic::amdgcn_permlane64, V);
427     return buildNonAtomicBinOp(B, Op, V, Permlane64Call);
428   }
429 
430   // Pick an arbitrary lane from 0..31 and an arbitrary lane from 32..63 and
431   // combine them with a scalar operation.
432   Function *ReadLane = Intrinsic::getOrInsertDeclaration(
433       M, Intrinsic::amdgcn_readlane, AtomicTy);
434   Value *Lane0 = B.CreateCall(ReadLane, {V, B.getInt32(0)});
435   Value *Lane32 = B.CreateCall(ReadLane, {V, B.getInt32(32)});
436   return buildNonAtomicBinOp(B, Op, Lane0, Lane32);
437 }
438 
439 // Use the builder to create an inclusive scan of V across the wavefront, with
440 // all lanes active.
441 Value *AMDGPUAtomicOptimizerImpl::buildScan(IRBuilder<> &B,
442                                             AtomicRMWInst::BinOp Op, Value *V,
443                                             Value *Identity) const {
444   Type *AtomicTy = V->getType();
445   Module *M = B.GetInsertBlock()->getModule();
446   Function *UpdateDPP = Intrinsic::getOrInsertDeclaration(
447       M, Intrinsic::amdgcn_update_dpp, AtomicTy);
448 
449   for (unsigned Idx = 0; Idx < 4; Idx++) {
450     V = buildNonAtomicBinOp(
451         B, Op, V,
452         B.CreateCall(UpdateDPP,
453                      {Identity, V, B.getInt32(DPP::ROW_SHR0 | 1 << Idx),
454                       B.getInt32(0xf), B.getInt32(0xf), B.getFalse()}));
455   }
456   if (ST.hasDPPBroadcasts()) {
457     // GFX9 has DPP row broadcast operations.
458     V = buildNonAtomicBinOp(
459         B, Op, V,
460         B.CreateCall(UpdateDPP,
461                      {Identity, V, B.getInt32(DPP::BCAST15), B.getInt32(0xa),
462                       B.getInt32(0xf), B.getFalse()}));
463     V = buildNonAtomicBinOp(
464         B, Op, V,
465         B.CreateCall(UpdateDPP,
466                      {Identity, V, B.getInt32(DPP::BCAST31), B.getInt32(0xc),
467                       B.getInt32(0xf), B.getFalse()}));
468   } else {
469     // On GFX10 all DPP operations are confined to a single row. To get cross-
470     // row operations we have to use permlane or readlane.
471 
472     // Combine lane 15 into lanes 16..31 (and, for wave 64, lane 47 into lanes
473     // 48..63).
474     assert(ST.hasPermLaneX16());
475     Value *PermX =
476         B.CreateIntrinsic(AtomicTy, Intrinsic::amdgcn_permlanex16,
477                           {PoisonValue::get(AtomicTy), V, B.getInt32(-1),
478                            B.getInt32(-1), B.getFalse(), B.getFalse()});
479 
480     Value *UpdateDPPCall = B.CreateCall(
481         UpdateDPP, {Identity, PermX, B.getInt32(DPP::QUAD_PERM_ID),
482                     B.getInt32(0xa), B.getInt32(0xf), B.getFalse()});
483     V = buildNonAtomicBinOp(B, Op, V, UpdateDPPCall);
484 
485     if (!ST.isWave32()) {
486       // Combine lane 31 into lanes 32..63.
487       Value *const Lane31 = B.CreateIntrinsic(
488           AtomicTy, Intrinsic::amdgcn_readlane, {V, B.getInt32(31)});
489 
490       Value *UpdateDPPCall = B.CreateCall(
491           UpdateDPP, {Identity, Lane31, B.getInt32(DPP::QUAD_PERM_ID),
492                       B.getInt32(0xc), B.getInt32(0xf), B.getFalse()});
493 
494       V = buildNonAtomicBinOp(B, Op, V, UpdateDPPCall);
495     }
496   }
497   return V;
498 }
499 
500 // Use the builder to create a shift right of V across the wavefront, with all
501 // lanes active, to turn an inclusive scan into an exclusive scan.
502 Value *AMDGPUAtomicOptimizerImpl::buildShiftRight(IRBuilder<> &B, Value *V,
503                                                   Value *Identity) const {
504   Type *AtomicTy = V->getType();
505   Module *M = B.GetInsertBlock()->getModule();
506   Function *UpdateDPP = Intrinsic::getOrInsertDeclaration(
507       M, Intrinsic::amdgcn_update_dpp, AtomicTy);
508   if (ST.hasDPPWavefrontShifts()) {
509     // GFX9 has DPP wavefront shift operations.
510     V = B.CreateCall(UpdateDPP,
511                      {Identity, V, B.getInt32(DPP::WAVE_SHR1), B.getInt32(0xf),
512                       B.getInt32(0xf), B.getFalse()});
513   } else {
514     Function *ReadLane = Intrinsic::getOrInsertDeclaration(
515         M, Intrinsic::amdgcn_readlane, AtomicTy);
516     Function *WriteLane = Intrinsic::getOrInsertDeclaration(
517         M, Intrinsic::amdgcn_writelane, AtomicTy);
518 
519     // On GFX10 all DPP operations are confined to a single row. To get cross-
520     // row operations we have to use permlane or readlane.
521     Value *Old = V;
522     V = B.CreateCall(UpdateDPP,
523                      {Identity, V, B.getInt32(DPP::ROW_SHR0 + 1),
524                       B.getInt32(0xf), B.getInt32(0xf), B.getFalse()});
525 
526     // Copy the old lane 15 to the new lane 16.
527     V = B.CreateCall(WriteLane, {B.CreateCall(ReadLane, {Old, B.getInt32(15)}),
528                                  B.getInt32(16), V});
529 
530     if (!ST.isWave32()) {
531       // Copy the old lane 31 to the new lane 32.
532       V = B.CreateCall(
533           WriteLane,
534           {B.CreateCall(ReadLane, {Old, B.getInt32(31)}), B.getInt32(32), V});
535 
536       // Copy the old lane 47 to the new lane 48.
537       V = B.CreateCall(
538           WriteLane,
539           {B.CreateCall(ReadLane, {Old, B.getInt32(47)}), B.getInt32(48), V});
540     }
541   }
542 
543   return V;
544 }
545 
546 // Use the builder to create an exclusive scan and compute the final reduced
547 // value using an iterative approach. This provides an alternative
548 // implementation to DPP which uses WMM for scan computations. This API iterate
549 // over active lanes to read, compute and update the value using
550 // readlane and writelane intrinsics.
551 std::pair<Value *, Value *> AMDGPUAtomicOptimizerImpl::buildScanIteratively(
552     IRBuilder<> &B, AtomicRMWInst::BinOp Op, Value *const Identity, Value *V,
553     Instruction &I, BasicBlock *ComputeLoop, BasicBlock *ComputeEnd) const {
554   auto *Ty = I.getType();
555   auto *WaveTy = B.getIntNTy(ST.getWavefrontSize());
556   auto *EntryBB = I.getParent();
557   auto NeedResult = !I.use_empty();
558 
559   auto *Ballot =
560       B.CreateIntrinsic(Intrinsic::amdgcn_ballot, WaveTy, B.getTrue());
561 
562   // Start inserting instructions for ComputeLoop block
563   B.SetInsertPoint(ComputeLoop);
564   // Phi nodes for Accumulator, Scan results destination, and Active Lanes
565   auto *Accumulator = B.CreatePHI(Ty, 2, "Accumulator");
566   Accumulator->addIncoming(Identity, EntryBB);
567   PHINode *OldValuePhi = nullptr;
568   if (NeedResult) {
569     OldValuePhi = B.CreatePHI(Ty, 2, "OldValuePhi");
570     OldValuePhi->addIncoming(PoisonValue::get(Ty), EntryBB);
571   }
572   auto *ActiveBits = B.CreatePHI(WaveTy, 2, "ActiveBits");
573   ActiveBits->addIncoming(Ballot, EntryBB);
574 
575   // Use llvm.cttz intrinsic to find the lowest remaining active lane.
576   auto *FF1 =
577       B.CreateIntrinsic(Intrinsic::cttz, WaveTy, {ActiveBits, B.getTrue()});
578 
579   auto *LaneIdxInt = B.CreateTrunc(FF1, B.getInt32Ty());
580 
581   // Get the value required for atomic operation
582   Value *LaneValue = B.CreateIntrinsic(V->getType(), Intrinsic::amdgcn_readlane,
583                                        {V, LaneIdxInt});
584 
585   // Perform writelane if intermediate scan results are required later in the
586   // kernel computations
587   Value *OldValue = nullptr;
588   if (NeedResult) {
589     OldValue = B.CreateIntrinsic(V->getType(), Intrinsic::amdgcn_writelane,
590                                  {Accumulator, LaneIdxInt, OldValuePhi});
591     OldValuePhi->addIncoming(OldValue, ComputeLoop);
592   }
593 
594   // Accumulate the results
595   auto *NewAccumulator = buildNonAtomicBinOp(B, Op, Accumulator, LaneValue);
596   Accumulator->addIncoming(NewAccumulator, ComputeLoop);
597 
598   // Set bit to zero of current active lane so that for next iteration llvm.cttz
599   // return the next active lane
600   auto *Mask = B.CreateShl(ConstantInt::get(WaveTy, 1), FF1);
601 
602   auto *InverseMask = B.CreateXor(Mask, ConstantInt::get(WaveTy, -1));
603   auto *NewActiveBits = B.CreateAnd(ActiveBits, InverseMask);
604   ActiveBits->addIncoming(NewActiveBits, ComputeLoop);
605 
606   // Branch out of the loop when all lanes are processed.
607   auto *IsEnd = B.CreateICmpEQ(NewActiveBits, ConstantInt::get(WaveTy, 0));
608   B.CreateCondBr(IsEnd, ComputeEnd, ComputeLoop);
609 
610   B.SetInsertPoint(ComputeEnd);
611 
612   return {OldValue, NewAccumulator};
613 }
614 
615 static Constant *getIdentityValueForAtomicOp(Type *const Ty,
616                                              AtomicRMWInst::BinOp Op) {
617   LLVMContext &C = Ty->getContext();
618   const unsigned BitWidth = Ty->getPrimitiveSizeInBits();
619   switch (Op) {
620   default:
621     llvm_unreachable("Unhandled atomic op");
622   case AtomicRMWInst::Add:
623   case AtomicRMWInst::Sub:
624   case AtomicRMWInst::Or:
625   case AtomicRMWInst::Xor:
626   case AtomicRMWInst::UMax:
627     return ConstantInt::get(C, APInt::getMinValue(BitWidth));
628   case AtomicRMWInst::And:
629   case AtomicRMWInst::UMin:
630     return ConstantInt::get(C, APInt::getMaxValue(BitWidth));
631   case AtomicRMWInst::Max:
632     return ConstantInt::get(C, APInt::getSignedMinValue(BitWidth));
633   case AtomicRMWInst::Min:
634     return ConstantInt::get(C, APInt::getSignedMaxValue(BitWidth));
635   case AtomicRMWInst::FAdd:
636     return ConstantFP::get(C, APFloat::getZero(Ty->getFltSemantics(), true));
637   case AtomicRMWInst::FSub:
638     return ConstantFP::get(C, APFloat::getZero(Ty->getFltSemantics(), false));
639   case AtomicRMWInst::FMin:
640   case AtomicRMWInst::FMax:
641     // FIXME: atomicrmw fmax/fmin behave like llvm.maxnum/minnum so NaN is the
642     // closest thing they have to an identity, but it still does not preserve
643     // the difference between quiet and signaling NaNs or NaNs with different
644     // payloads.
645     return ConstantFP::get(C, APFloat::getNaN(Ty->getFltSemantics()));
646   }
647 }
648 
649 static Value *buildMul(IRBuilder<> &B, Value *LHS, Value *RHS) {
650   const ConstantInt *CI = dyn_cast<ConstantInt>(LHS);
651   return (CI && CI->isOne()) ? RHS : B.CreateMul(LHS, RHS);
652 }
653 
654 void AMDGPUAtomicOptimizerImpl::optimizeAtomic(Instruction &I,
655                                                AtomicRMWInst::BinOp Op,
656                                                unsigned ValIdx,
657                                                bool ValDivergent) const {
658   // Start building just before the instruction.
659   IRBuilder<> B(&I);
660 
661   if (AtomicRMWInst::isFPOperation(Op)) {
662     B.setIsFPConstrained(I.getFunction()->hasFnAttribute(Attribute::StrictFP));
663   }
664 
665   // If we are in a pixel shader, because of how we have to mask out helper
666   // lane invocations, we need to record the entry and exit BB's.
667   BasicBlock *PixelEntryBB = nullptr;
668   BasicBlock *PixelExitBB = nullptr;
669 
670   // If we're optimizing an atomic within a pixel shader, we need to wrap the
671   // entire atomic operation in a helper-lane check. We do not want any helper
672   // lanes that are around only for the purposes of derivatives to take part
673   // in any cross-lane communication, and we use a branch on whether the lane is
674   // live to do this.
675   if (IsPixelShader) {
676     // Record I's original position as the entry block.
677     PixelEntryBB = I.getParent();
678 
679     Value *const Cond = B.CreateIntrinsic(Intrinsic::amdgcn_ps_live, {}, {});
680     Instruction *const NonHelperTerminator =
681         SplitBlockAndInsertIfThen(Cond, &I, false, nullptr, &DTU, nullptr);
682 
683     // Record I's new position as the exit block.
684     PixelExitBB = I.getParent();
685 
686     I.moveBefore(NonHelperTerminator->getIterator());
687     B.SetInsertPoint(&I);
688   }
689 
690   Type *const Ty = I.getType();
691   Type *Int32Ty = B.getInt32Ty();
692   bool isAtomicFloatingPointTy = Ty->isFloatingPointTy();
693   [[maybe_unused]] const unsigned TyBitWidth = DL.getTypeSizeInBits(Ty);
694 
695   // This is the value in the atomic operation we need to combine in order to
696   // reduce the number of atomic operations.
697   Value *V = I.getOperand(ValIdx);
698 
699   // We need to know how many lanes are active within the wavefront, and we do
700   // this by doing a ballot of active lanes.
701   Type *const WaveTy = B.getIntNTy(ST.getWavefrontSize());
702   CallInst *const Ballot =
703       B.CreateIntrinsic(Intrinsic::amdgcn_ballot, WaveTy, B.getTrue());
704 
705   // We need to know how many lanes are active within the wavefront that are
706   // below us. If we counted each lane linearly starting from 0, a lane is
707   // below us only if its associated index was less than ours. We do this by
708   // using the mbcnt intrinsic.
709   Value *Mbcnt;
710   if (ST.isWave32()) {
711     Mbcnt = B.CreateIntrinsic(Intrinsic::amdgcn_mbcnt_lo, {},
712                               {Ballot, B.getInt32(0)});
713   } else {
714     Value *const ExtractLo = B.CreateTrunc(Ballot, Int32Ty);
715     Value *const ExtractHi = B.CreateTrunc(B.CreateLShr(Ballot, 32), Int32Ty);
716     Mbcnt = B.CreateIntrinsic(Intrinsic::amdgcn_mbcnt_lo, {},
717                               {ExtractLo, B.getInt32(0)});
718     Mbcnt =
719         B.CreateIntrinsic(Intrinsic::amdgcn_mbcnt_hi, {}, {ExtractHi, Mbcnt});
720   }
721 
722   Function *F = I.getFunction();
723   LLVMContext &C = F->getContext();
724 
725   // For atomic sub, perform scan with add operation and allow one lane to
726   // subtract the reduced value later.
727   AtomicRMWInst::BinOp ScanOp = Op;
728   if (Op == AtomicRMWInst::Sub) {
729     ScanOp = AtomicRMWInst::Add;
730   } else if (Op == AtomicRMWInst::FSub) {
731     ScanOp = AtomicRMWInst::FAdd;
732   }
733   Value *Identity = getIdentityValueForAtomicOp(Ty, ScanOp);
734 
735   Value *ExclScan = nullptr;
736   Value *NewV = nullptr;
737 
738   const bool NeedResult = !I.use_empty();
739 
740   BasicBlock *ComputeLoop = nullptr;
741   BasicBlock *ComputeEnd = nullptr;
742   // If we have a divergent value in each lane, we need to combine the value
743   // using DPP.
744   if (ValDivergent) {
745     if (ScanImpl == ScanOptions::DPP) {
746       // First we need to set all inactive invocations to the identity value, so
747       // that they can correctly contribute to the final result.
748       NewV =
749           B.CreateIntrinsic(Intrinsic::amdgcn_set_inactive, Ty, {V, Identity});
750       if (!NeedResult && ST.hasPermLaneX16()) {
751         // On GFX10 the permlanex16 instruction helps us build a reduction
752         // without too many readlanes and writelanes, which are generally bad
753         // for performance.
754         NewV = buildReduction(B, ScanOp, NewV, Identity);
755       } else {
756         NewV = buildScan(B, ScanOp, NewV, Identity);
757         if (NeedResult)
758           ExclScan = buildShiftRight(B, NewV, Identity);
759         // Read the value from the last lane, which has accumulated the values
760         // of each active lane in the wavefront. This will be our new value
761         // which we will provide to the atomic operation.
762         Value *const LastLaneIdx = B.getInt32(ST.getWavefrontSize() - 1);
763         NewV = B.CreateIntrinsic(Ty, Intrinsic::amdgcn_readlane,
764                                  {NewV, LastLaneIdx});
765       }
766       // Finally mark the readlanes in the WWM section.
767       NewV = B.CreateIntrinsic(Intrinsic::amdgcn_strict_wwm, Ty, NewV);
768     } else if (ScanImpl == ScanOptions::Iterative) {
769       // Alternative implementation for scan
770       ComputeLoop = BasicBlock::Create(C, "ComputeLoop", F);
771       ComputeEnd = BasicBlock::Create(C, "ComputeEnd", F);
772       std::tie(ExclScan, NewV) = buildScanIteratively(B, ScanOp, Identity, V, I,
773                                                       ComputeLoop, ComputeEnd);
774     } else {
775       llvm_unreachable("Atomic Optimzer is disabled for None strategy");
776     }
777   } else {
778     switch (Op) {
779     default:
780       llvm_unreachable("Unhandled atomic op");
781 
782     case AtomicRMWInst::Add:
783     case AtomicRMWInst::Sub: {
784       // The new value we will be contributing to the atomic operation is the
785       // old value times the number of active lanes.
786       Value *const Ctpop = B.CreateIntCast(
787           B.CreateUnaryIntrinsic(Intrinsic::ctpop, Ballot), Ty, false);
788       NewV = buildMul(B, V, Ctpop);
789       break;
790     }
791     case AtomicRMWInst::FAdd:
792     case AtomicRMWInst::FSub: {
793       Value *const Ctpop = B.CreateIntCast(
794           B.CreateUnaryIntrinsic(Intrinsic::ctpop, Ballot), Int32Ty, false);
795       Value *const CtpopFP = B.CreateUIToFP(Ctpop, Ty);
796       NewV = B.CreateFMul(V, CtpopFP);
797       break;
798     }
799     case AtomicRMWInst::And:
800     case AtomicRMWInst::Or:
801     case AtomicRMWInst::Max:
802     case AtomicRMWInst::Min:
803     case AtomicRMWInst::UMax:
804     case AtomicRMWInst::UMin:
805     case AtomicRMWInst::FMin:
806     case AtomicRMWInst::FMax:
807       // These operations with a uniform value are idempotent: doing the atomic
808       // operation multiple times has the same effect as doing it once.
809       NewV = V;
810       break;
811 
812     case AtomicRMWInst::Xor:
813       // The new value we will be contributing to the atomic operation is the
814       // old value times the parity of the number of active lanes.
815       Value *const Ctpop = B.CreateIntCast(
816           B.CreateUnaryIntrinsic(Intrinsic::ctpop, Ballot), Ty, false);
817       NewV = buildMul(B, V, B.CreateAnd(Ctpop, 1));
818       break;
819     }
820   }
821 
822   // We only want a single lane to enter our new control flow, and we do this
823   // by checking if there are any active lanes below us. Only one lane will
824   // have 0 active lanes below us, so that will be the only one to progress.
825   Value *const Cond = B.CreateICmpEQ(Mbcnt, B.getInt32(0));
826 
827   // Store I's original basic block before we split the block.
828   BasicBlock *const OriginalBB = I.getParent();
829 
830   // We need to introduce some new control flow to force a single lane to be
831   // active. We do this by splitting I's basic block at I, and introducing the
832   // new block such that:
833   // entry --> single_lane -\
834   //       \------------------> exit
835   Instruction *const SingleLaneTerminator =
836       SplitBlockAndInsertIfThen(Cond, &I, false, nullptr, &DTU, nullptr);
837 
838   // At this point, we have split the I's block to allow one lane in wavefront
839   // to update the precomputed reduced value. Also, completed the codegen for
840   // new control flow i.e. iterative loop which perform reduction and scan using
841   // ComputeLoop and ComputeEnd.
842   // For the new control flow, we need to move branch instruction i.e.
843   // terminator created during SplitBlockAndInsertIfThen from I's block to
844   // ComputeEnd block. We also need to set up predecessor to next block when
845   // single lane done updating the final reduced value.
846   BasicBlock *Predecessor = nullptr;
847   if (ValDivergent && ScanImpl == ScanOptions::Iterative) {
848     // Move terminator from I's block to ComputeEnd block.
849     //
850     // OriginalBB is known to have a branch as terminator because
851     // SplitBlockAndInsertIfThen will have inserted one.
852     BranchInst *Terminator = cast<BranchInst>(OriginalBB->getTerminator());
853     B.SetInsertPoint(ComputeEnd);
854     Terminator->removeFromParent();
855     B.Insert(Terminator);
856 
857     // Branch to ComputeLoop Block unconditionally from the I's block for
858     // iterative approach.
859     B.SetInsertPoint(OriginalBB);
860     B.CreateBr(ComputeLoop);
861 
862     // Update the dominator tree for new control flow.
863     SmallVector<DominatorTree::UpdateType, 6> DomTreeUpdates(
864         {{DominatorTree::Insert, OriginalBB, ComputeLoop},
865          {DominatorTree::Insert, ComputeLoop, ComputeEnd}});
866 
867     // We're moving the terminator from EntryBB to ComputeEnd, make sure we move
868     // the DT edges as well.
869     for (auto *Succ : Terminator->successors()) {
870       DomTreeUpdates.push_back({DominatorTree::Insert, ComputeEnd, Succ});
871       DomTreeUpdates.push_back({DominatorTree::Delete, OriginalBB, Succ});
872     }
873 
874     DTU.applyUpdates(DomTreeUpdates);
875 
876     Predecessor = ComputeEnd;
877   } else {
878     Predecessor = OriginalBB;
879   }
880   // Move the IR builder into single_lane next.
881   B.SetInsertPoint(SingleLaneTerminator);
882 
883   // Clone the original atomic operation into single lane, replacing the
884   // original value with our newly created one.
885   Instruction *const NewI = I.clone();
886   B.Insert(NewI);
887   NewI->setOperand(ValIdx, NewV);
888 
889   // Move the IR builder into exit next, and start inserting just before the
890   // original instruction.
891   B.SetInsertPoint(&I);
892 
893   if (NeedResult) {
894     // Create a PHI node to get our new atomic result into the exit block.
895     PHINode *const PHI = B.CreatePHI(Ty, 2);
896     PHI->addIncoming(PoisonValue::get(Ty), Predecessor);
897     PHI->addIncoming(NewI, SingleLaneTerminator->getParent());
898 
899     // We need to broadcast the value who was the lowest active lane (the first
900     // lane) to all other lanes in the wavefront.
901     Value *BroadcastI = nullptr;
902     BroadcastI = B.CreateIntrinsic(Ty, Intrinsic::amdgcn_readfirstlane, PHI);
903 
904     // Now that we have the result of our single atomic operation, we need to
905     // get our individual lane's slice into the result. We use the lane offset
906     // we previously calculated combined with the atomic result value we got
907     // from the first lane, to get our lane's index into the atomic result.
908     Value *LaneOffset = nullptr;
909     if (ValDivergent) {
910       if (ScanImpl == ScanOptions::DPP) {
911         LaneOffset =
912             B.CreateIntrinsic(Intrinsic::amdgcn_strict_wwm, Ty, ExclScan);
913       } else if (ScanImpl == ScanOptions::Iterative) {
914         LaneOffset = ExclScan;
915       } else {
916         llvm_unreachable("Atomic Optimzer is disabled for None strategy");
917       }
918     } else {
919       Mbcnt = isAtomicFloatingPointTy ? B.CreateUIToFP(Mbcnt, Ty)
920                                       : B.CreateIntCast(Mbcnt, Ty, false);
921       switch (Op) {
922       default:
923         llvm_unreachable("Unhandled atomic op");
924       case AtomicRMWInst::Add:
925       case AtomicRMWInst::Sub:
926         LaneOffset = buildMul(B, V, Mbcnt);
927         break;
928       case AtomicRMWInst::And:
929       case AtomicRMWInst::Or:
930       case AtomicRMWInst::Max:
931       case AtomicRMWInst::Min:
932       case AtomicRMWInst::UMax:
933       case AtomicRMWInst::UMin:
934       case AtomicRMWInst::FMin:
935       case AtomicRMWInst::FMax:
936         LaneOffset = B.CreateSelect(Cond, Identity, V);
937         break;
938       case AtomicRMWInst::Xor:
939         LaneOffset = buildMul(B, V, B.CreateAnd(Mbcnt, 1));
940         break;
941       case AtomicRMWInst::FAdd:
942       case AtomicRMWInst::FSub: {
943         LaneOffset = B.CreateFMul(V, Mbcnt);
944         break;
945       }
946       }
947     }
948     Value *Result = buildNonAtomicBinOp(B, Op, BroadcastI, LaneOffset);
949     if (isAtomicFloatingPointTy) {
950       // For fadd/fsub the first active lane of LaneOffset should be the
951       // identity (-0.0 for fadd or +0.0 for fsub) but the value we calculated
952       // is V * +0.0 which might have the wrong sign or might be nan (if V is
953       // inf or nan).
954       //
955       // For all floating point ops if the in-memory value was a nan then the
956       // binop we just built might have quieted it or changed its payload.
957       //
958       // Correct all these problems by using BroadcastI as the result in the
959       // first active lane.
960       Result = B.CreateSelect(Cond, BroadcastI, Result);
961     }
962 
963     if (IsPixelShader) {
964       // Need a final PHI to reconverge to above the helper lane branch mask.
965       B.SetInsertPoint(PixelExitBB, PixelExitBB->getFirstNonPHIIt());
966 
967       PHINode *const PHI = B.CreatePHI(Ty, 2);
968       PHI->addIncoming(PoisonValue::get(Ty), PixelEntryBB);
969       PHI->addIncoming(Result, I.getParent());
970       I.replaceAllUsesWith(PHI);
971     } else {
972       // Replace the original atomic instruction with the new one.
973       I.replaceAllUsesWith(Result);
974     }
975   }
976 
977   // And delete the original.
978   I.eraseFromParent();
979 }
980 
981 INITIALIZE_PASS_BEGIN(AMDGPUAtomicOptimizer, DEBUG_TYPE,
982                       "AMDGPU atomic optimizations", false, false)
983 INITIALIZE_PASS_DEPENDENCY(UniformityInfoWrapperPass)
984 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
985 INITIALIZE_PASS_END(AMDGPUAtomicOptimizer, DEBUG_TYPE,
986                     "AMDGPU atomic optimizations", false, false)
987 
988 FunctionPass *llvm::createAMDGPUAtomicOptimizerPass(ScanOptions ScanStrategy) {
989   return new AMDGPUAtomicOptimizer(ScanStrategy);
990 }
991