xref: /llvm-project/llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp (revision 7b9d73c2f90c0ed8497339a16fc39785349d9610)
1 //===-- AMDGPULowerModuleLDSPass.cpp ------------------------------*- C++ -*-=//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass eliminates local data store, LDS, uses from non-kernel functions.
10 // LDS is contiguous memory allocated per kernel execution.
11 //
12 // Background.
13 //
14 // The programming model is global variables, or equivalently function local
15 // static variables, accessible from kernels or other functions. For uses from
16 // kernels this is straightforward - assign an integer to the kernel for the
17 // memory required by all the variables combined, allocate them within that.
18 // For uses from functions there are performance tradeoffs to choose between.
19 //
20 // This model means the GPU runtime can specify the amount of memory allocated.
21 // If this is more than the kernel assumed, the excess can be made available
22 // using a language specific feature, which IR represents as a variable with
23 // no initializer. This feature is referred to here as "Dynamic LDS" and is
24 // lowered slightly differently to the normal case.
25 //
26 // Consequences of this GPU feature:
27 // - memory is limited and exceeding it halts compilation
28 // - a global accessed by one kernel exists independent of other kernels
29 // - a global exists independent of simultaneous execution of the same kernel
30 // - the address of the global may be different from different kernels as they
31 //   do not alias, which permits only allocating variables they use
32 // - if the address is allowed to differ, functions need help to find it
33 //
34 // Uses from kernels are implemented here by grouping them in a per-kernel
35 // struct instance. This duplicates the variables, accurately modelling their
36 // aliasing properties relative to a single global representation. It also
37 // permits control over alignment via padding.
38 //
39 // Uses from functions are more complicated and the primary purpose of this
40 // IR pass. Several different lowering are chosen between to meet requirements
41 // to avoid allocating any LDS where it is not necessary, as that impacts
42 // occupancy and may fail the compilation, while not imposing overhead on a
43 // feature whose primary advantage over global memory is performance. The basic
44 // design goal is to avoid one kernel imposing overhead on another.
45 //
46 // Implementation.
47 //
48 // LDS variables with constant annotation or non-undef initializer are passed
49 // through unchanged for simplification or error diagnostics in later passes.
50 // Non-undef initializers are not yet implemented for LDS.
51 //
52 // LDS variables that are always allocated at the same address can be found
53 // by lookup at that address. Otherwise runtime information/cost is required.
54 //
55 // The simplest strategy possible is to group all LDS variables in a single
56 // struct and allocate that struct in every kernel such that the original
57 // variables are always at the same address. LDS is however a limited resource
58 // so this strategy is unusable in practice. It is not implemented here.
59 //
60 // Strategy | Precise allocation | Zero runtime cost | General purpose |
61 //  --------+--------------------+-------------------+-----------------+
62 //   Module |                 No |               Yes |             Yes |
63 //    Table |                Yes |                No |             Yes |
64 //   Kernel |                Yes |               Yes |              No |
65 //   Hybrid |                Yes |           Partial |             Yes |
66 //
67 // "Module" spends LDS memory to save cycles. "Table" spends cycles and global
68 // memory to save LDS. "Kernel" is as fast as kernel allocation but only works
69 // for variables that are known reachable from a single kernel. "Hybrid" picks
70 // between all three. When forced to choose between LDS and cycles we minimise
71 // LDS use.
72 
73 // The "module" lowering implemented here finds LDS variables which are used by
74 // non-kernel functions and creates a new struct with a field for each of those
75 // LDS variables. Variables that are only used from kernels are excluded.
76 //
77 // The "table" lowering implemented here has three components.
78 // First kernels are assigned a unique integer identifier which is available in
79 // functions it calls through the intrinsic amdgcn_lds_kernel_id. The integer
80 // is passed through a specific SGPR, thus works with indirect calls.
81 // Second, each kernel allocates LDS variables independent of other kernels and
82 // writes the addresses it chose for each variable into an array in consistent
83 // order. If the kernel does not allocate a given variable, it writes undef to
84 // the corresponding array location. These arrays are written to a constant
85 // table in the order matching the kernel unique integer identifier.
86 // Third, uses from non-kernel functions are replaced with a table lookup using
87 // the intrinsic function to find the address of the variable.
88 //
89 // "Kernel" lowering is only applicable for variables that are unambiguously
90 // reachable from exactly one kernel. For those cases, accesses to the variable
91 // can be lowered to ConstantExpr address of a struct instance specific to that
92 // one kernel. This is zero cost in space and in compute. It will raise a fatal
93 // error on any variable that might be reachable from multiple kernels and is
94 // thus most easily used as part of the hybrid lowering strategy.
95 //
96 // Hybrid lowering is a mixture of the above. It uses the zero cost kernel
97 // lowering where it can. It lowers the variable accessed by the greatest
98 // number of kernels using the module strategy as that is free for the first
99 // variable. Any futher variables that can be lowered with the module strategy
100 // without incurring LDS memory overhead are. The remaining ones are lowered
101 // via table.
102 //
103 // Consequences
104 // - No heuristics or user controlled magic numbers, hybrid is the right choice
105 // - Kernels that don't use functions (or have had them all inlined) are not
106 //   affected by any lowering for kernels that do.
107 // - Kernels that don't make indirect function calls are not affected by those
108 //   that do.
109 // - Variables which are used by lots of kernels, e.g. those injected by a
110 //   language runtime in most kernels, are expected to have no overhead
111 // - Implementations that instantiate templates per-kernel where those templates
112 //   use LDS are expected to hit the "Kernel" lowering strategy
113 // - The runtime properties impose a cost in compiler implementation complexity
114 //
115 // Dynamic LDS implementation
116 // Dynamic LDS is lowered similarly to the "table" strategy above and uses the
117 // same intrinsic to identify which kernel is at the root of the dynamic call
118 // graph. This relies on the specified behaviour that all dynamic LDS variables
119 // alias one another, i.e. are at the same address, with respect to a given
120 // kernel. Therefore this pass creates new dynamic LDS variables for each kernel
121 // that allocates any dynamic LDS and builds a table of addresses out of those.
122 // The AMDGPUPromoteAlloca pass skips kernels that use dynamic LDS.
123 // The corresponding optimisation for "kernel" lowering where the table lookup
124 // is elided is not implemented.
125 //
126 //
127 // Implementation notes / limitations
128 // A single LDS global variable represents an instance per kernel that can reach
129 // said variables. This pass essentially specialises said variables per kernel.
130 // Handling ConstantExpr during the pass complicated this significantly so now
131 // all ConstantExpr uses of LDS variables are expanded to instructions. This
132 // may need amending when implementing non-undef initialisers.
133 //
134 // Lowering is split between this IR pass and the back end. This pass chooses
135 // where given variables should be allocated and marks them with metadata,
136 // MD_absolute_symbol. The backend places the variables in coincidentally the
137 // same location and raises a fatal error if something has gone awry. This works
138 // in practice because the only pass between this one and the backend that
139 // changes LDS is PromoteAlloca and the changes it makes do not conflict.
140 //
141 // Addresses are written to constant global arrays based on the same metadata.
142 //
143 // The backend lowers LDS variables in the order of traversal of the function.
144 // This is at odds with the deterministic layout required. The workaround is to
145 // allocate the fixed-address variables immediately upon starting the function
146 // where they can be placed as intended. This requires a means of mapping from
147 // the function to the variables that it allocates. For the module scope lds,
148 // this is via metadata indicating whether the variable is not required. If a
149 // pass deletes that metadata, a fatal error on disagreement with the absolute
150 // symbol metadata will occur. For kernel scope and dynamic, this is by _name_
151 // correspondence between the function and the variable. It requires the
152 // kernel to have a name (which is only a limitation for tests in practice) and
153 // for nothing to rename the corresponding symbols. This is a hazard if the pass
154 // is run multiple times during debugging. Alternative schemes considered all
155 // involve bespoke metadata.
156 //
157 // If the name correspondence can be replaced, multiple distinct kernels that
158 // have the same memory layout can map to the same kernel id (as the address
159 // itself is handled by the absolute symbol metadata) and that will allow more
160 // uses of the "kernel" style faster lowering and reduce the size of the lookup
161 // tables.
162 //
163 // There is a test that checks this does not fire for a graphics shader. This
164 // lowering is expected to work for graphics if the isKernel test is changed.
165 //
166 // The current markUsedByKernel is sufficient for PromoteAlloca but is elided
167 // before codegen. Replacing this with an equivalent intrinsic which lasts until
168 // shortly after the machine function lowering of LDS would help break the name
169 // mapping. The other part needed is probably to amend PromoteAlloca to embed
170 // the LDS variables it creates in the same struct created here. That avoids the
171 // current hazard where a PromoteAlloca LDS variable might be allocated before
172 // the kernel scope (and thus error on the address check). Given a new invariant
173 // that no LDS variables exist outside of the structs managed here, and an
174 // intrinsic that lasts until after the LDS frame lowering, it should be
175 // possible to drop the name mapping and fold equivalent memory layouts.
176 //
177 //===----------------------------------------------------------------------===//
178 
179 #include "AMDGPU.h"
180 #include "AMDGPUTargetMachine.h"
181 #include "Utils/AMDGPUBaseInfo.h"
182 #include "Utils/AMDGPUMemoryUtils.h"
183 #include "llvm/ADT/BitVector.h"
184 #include "llvm/ADT/DenseMap.h"
185 #include "llvm/ADT/DenseSet.h"
186 #include "llvm/ADT/STLExtras.h"
187 #include "llvm/ADT/SetOperations.h"
188 #include "llvm/ADT/SetVector.h"
189 #include "llvm/Analysis/CallGraph.h"
190 #include "llvm/CodeGen/TargetPassConfig.h"
191 #include "llvm/IR/Constants.h"
192 #include "llvm/IR/DerivedTypes.h"
193 #include "llvm/IR/IRBuilder.h"
194 #include "llvm/IR/InlineAsm.h"
195 #include "llvm/IR/Instructions.h"
196 #include "llvm/IR/IntrinsicsAMDGPU.h"
197 #include "llvm/IR/MDBuilder.h"
198 #include "llvm/IR/ReplaceConstant.h"
199 #include "llvm/InitializePasses.h"
200 #include "llvm/Pass.h"
201 #include "llvm/Support/CommandLine.h"
202 #include "llvm/Support/Debug.h"
203 #include "llvm/Support/Format.h"
204 #include "llvm/Support/OptimizedStructLayout.h"
205 #include "llvm/Support/raw_ostream.h"
206 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
207 #include "llvm/Transforms/Utils/ModuleUtils.h"
208 
209 #include <tuple>
210 #include <vector>
211 
212 #include <cstdio>
213 
214 #define DEBUG_TYPE "amdgpu-lower-module-lds"
215 
216 using namespace llvm;
217 
218 namespace {
219 
220 cl::opt<bool> SuperAlignLDSGlobals(
221     "amdgpu-super-align-lds-globals",
222     cl::desc("Increase alignment of LDS if it is not on align boundary"),
223     cl::init(true), cl::Hidden);
224 
225 enum class LoweringKind { module, table, kernel, hybrid };
226 cl::opt<LoweringKind> LoweringKindLoc(
227     "amdgpu-lower-module-lds-strategy",
228     cl::desc("Specify lowering strategy for function LDS access:"), cl::Hidden,
229     cl::init(LoweringKind::hybrid),
230     cl::values(
231         clEnumValN(LoweringKind::table, "table", "Lower via table lookup"),
232         clEnumValN(LoweringKind::module, "module", "Lower via module struct"),
233         clEnumValN(
234             LoweringKind::kernel, "kernel",
235             "Lower variables reachable from one kernel, otherwise abort"),
236         clEnumValN(LoweringKind::hybrid, "hybrid",
237                    "Lower via mixture of above strategies")));
238 
239 bool isKernelLDS(const Function *F) {
240   // Some weirdness here. AMDGPU::isKernelCC does not call into
241   // AMDGPU::isKernel with the calling conv, it instead calls into
242   // isModuleEntryFunction which returns true for more calling conventions
243   // than AMDGPU::isKernel does. There's a FIXME on AMDGPU::isKernel.
244   // There's also a test that checks that the LDS lowering does not hit on
245   // a graphics shader, denoted amdgpu_ps, so stay with the limited case.
246   // Putting LDS in the name of the function to draw attention to this.
247   return AMDGPU::isKernel(F->getCallingConv());
248 }
249 
250 template <typename T> std::vector<T> sortByName(std::vector<T> &&V) {
251   llvm::sort(V.begin(), V.end(), [](const auto *L, const auto *R) {
252     return L->getName() < R->getName();
253   });
254   return {std::move(V)};
255 }
256 
257 class AMDGPULowerModuleLDS {
258   const AMDGPUTargetMachine &TM;
259 
260   static void
261   removeLocalVarsFromUsedLists(Module &M,
262                                const DenseSet<GlobalVariable *> &LocalVars) {
263     // The verifier rejects used lists containing an inttoptr of a constant
264     // so remove the variables from these lists before replaceAllUsesWith
265     SmallPtrSet<Constant *, 8> LocalVarsSet;
266     for (GlobalVariable *LocalVar : LocalVars)
267       LocalVarsSet.insert(cast<Constant>(LocalVar->stripPointerCasts()));
268 
269     removeFromUsedLists(
270         M, [&LocalVarsSet](Constant *C) { return LocalVarsSet.count(C); });
271 
272     for (GlobalVariable *LocalVar : LocalVars)
273       LocalVar->removeDeadConstantUsers();
274   }
275 
276   static void markUsedByKernel(Function *Func, GlobalVariable *SGV) {
277     // The llvm.amdgcn.module.lds instance is implicitly used by all kernels
278     // that might call a function which accesses a field within it. This is
279     // presently approximated to 'all kernels' if there are any such functions
280     // in the module. This implicit use is redefined as an explicit use here so
281     // that later passes, specifically PromoteAlloca, account for the required
282     // memory without any knowledge of this transform.
283 
284     // An operand bundle on llvm.donothing works because the call instruction
285     // survives until after the last pass that needs to account for LDS. It is
286     // better than inline asm as the latter survives until the end of codegen. A
287     // totally robust solution would be a function with the same semantics as
288     // llvm.donothing that takes a pointer to the instance and is lowered to a
289     // no-op after LDS is allocated, but that is not presently necessary.
290 
291     // This intrinsic is eliminated shortly before instruction selection. It
292     // does not suffice to indicate to ISel that a given global which is not
293     // immediately used by the kernel must still be allocated by it. An
294     // equivalent target specific intrinsic which lasts until immediately after
295     // codegen would suffice for that, but one would still need to ensure that
296     // the variables are allocated in the anticpated order.
297     BasicBlock *Entry = &Func->getEntryBlock();
298     IRBuilder<> Builder(Entry, Entry->getFirstNonPHIIt());
299 
300     Function *Decl =
301         Intrinsic::getDeclaration(Func->getParent(), Intrinsic::donothing, {});
302 
303     Value *UseInstance[1] = {
304         Builder.CreateConstInBoundsGEP1_32(SGV->getValueType(), SGV, 0)};
305 
306     Builder.CreateCall(
307         Decl, {}, {OperandBundleDefT<Value *>("ExplicitUse", UseInstance)});
308   }
309 
310   static bool eliminateConstantExprUsesOfLDSFromAllInstructions(Module &M) {
311     // Constants are uniqued within LLVM. A ConstantExpr referring to a LDS
312     // global may have uses from multiple different functions as a result.
313     // This pass specialises LDS variables with respect to the kernel that
314     // allocates them.
315 
316     // This is semantically equivalent to (the unimplemented as slow):
317     // for (auto &F : M.functions())
318     //   for (auto &BB : F)
319     //     for (auto &I : BB)
320     //       for (Use &Op : I.operands())
321     //         if (constantExprUsesLDS(Op))
322     //           replaceConstantExprInFunction(I, Op);
323 
324     SmallVector<Constant *> LDSGlobals;
325     for (auto &GV : M.globals())
326       if (AMDGPU::isLDSVariableToLower(GV))
327         LDSGlobals.push_back(&GV);
328 
329     return convertUsersOfConstantsToInstructions(LDSGlobals);
330   }
331 
332 public:
333   AMDGPULowerModuleLDS(const AMDGPUTargetMachine &TM_) : TM(TM_) {}
334 
335   using FunctionVariableMap = DenseMap<Function *, DenseSet<GlobalVariable *>>;
336 
337   using VariableFunctionMap = DenseMap<GlobalVariable *, DenseSet<Function *>>;
338 
339   static void getUsesOfLDSByFunction(CallGraph const &CG, Module &M,
340                                      FunctionVariableMap &kernels,
341                                      FunctionVariableMap &functions) {
342 
343     // Get uses from the current function, excluding uses by called functions
344     // Two output variables to avoid walking the globals list twice
345     for (auto &GV : M.globals()) {
346       if (!AMDGPU::isLDSVariableToLower(GV)) {
347         continue;
348       }
349 
350       if (GV.isAbsoluteSymbolRef()) {
351         report_fatal_error(
352             "LDS variables with absolute addresses are unimplemented.");
353       }
354 
355       for (User *V : GV.users()) {
356         if (auto *I = dyn_cast<Instruction>(V)) {
357           Function *F = I->getFunction();
358           if (isKernelLDS(F)) {
359             kernels[F].insert(&GV);
360           } else {
361             functions[F].insert(&GV);
362           }
363         }
364       }
365     }
366   }
367 
368   struct LDSUsesInfoTy {
369     FunctionVariableMap direct_access;
370     FunctionVariableMap indirect_access;
371   };
372 
373   static LDSUsesInfoTy getTransitiveUsesOfLDS(CallGraph const &CG, Module &M) {
374 
375     FunctionVariableMap direct_map_kernel;
376     FunctionVariableMap direct_map_function;
377     getUsesOfLDSByFunction(CG, M, direct_map_kernel, direct_map_function);
378 
379     // Collect variables that are used by functions whose address has escaped
380     DenseSet<GlobalVariable *> VariablesReachableThroughFunctionPointer;
381     for (Function &F : M.functions()) {
382       if (!isKernelLDS(&F))
383         if (F.hasAddressTaken(nullptr,
384                               /* IgnoreCallbackUses */ false,
385                               /* IgnoreAssumeLikeCalls */ false,
386                               /* IgnoreLLVMUsed */ true,
387                               /* IgnoreArcAttachedCall */ false)) {
388           set_union(VariablesReachableThroughFunctionPointer,
389                     direct_map_function[&F]);
390         }
391     }
392 
393     auto functionMakesUnknownCall = [&](const Function *F) -> bool {
394       assert(!F->isDeclaration());
395       for (const CallGraphNode::CallRecord &R : *CG[F]) {
396         if (!R.second->getFunction()) {
397           return true;
398         }
399       }
400       return false;
401     };
402 
403     // Work out which variables are reachable through function calls
404     FunctionVariableMap transitive_map_function = direct_map_function;
405 
406     // If the function makes any unknown call, assume the worst case that it can
407     // access all variables accessed by functions whose address escaped
408     for (Function &F : M.functions()) {
409       if (!F.isDeclaration() && functionMakesUnknownCall(&F)) {
410         if (!isKernelLDS(&F)) {
411           set_union(transitive_map_function[&F],
412                     VariablesReachableThroughFunctionPointer);
413         }
414       }
415     }
416 
417     // Direct implementation of collecting all variables reachable from each
418     // function
419     for (Function &Func : M.functions()) {
420       if (Func.isDeclaration() || isKernelLDS(&Func))
421         continue;
422 
423       DenseSet<Function *> seen; // catches cycles
424       SmallVector<Function *, 4> wip{&Func};
425 
426       while (!wip.empty()) {
427         Function *F = wip.pop_back_val();
428 
429         // Can accelerate this by referring to transitive map for functions that
430         // have already been computed, with more care than this
431         set_union(transitive_map_function[&Func], direct_map_function[F]);
432 
433         for (const CallGraphNode::CallRecord &R : *CG[F]) {
434           Function *ith = R.second->getFunction();
435           if (ith) {
436             if (!seen.contains(ith)) {
437               seen.insert(ith);
438               wip.push_back(ith);
439             }
440           }
441         }
442       }
443     }
444 
445     // direct_map_kernel lists which variables are used by the kernel
446     // find the variables which are used through a function call
447     FunctionVariableMap indirect_map_kernel;
448 
449     for (Function &Func : M.functions()) {
450       if (Func.isDeclaration() || !isKernelLDS(&Func))
451         continue;
452 
453       for (const CallGraphNode::CallRecord &R : *CG[&Func]) {
454         Function *ith = R.second->getFunction();
455         if (ith) {
456           set_union(indirect_map_kernel[&Func], transitive_map_function[ith]);
457         } else {
458           set_union(indirect_map_kernel[&Func],
459                     VariablesReachableThroughFunctionPointer);
460         }
461       }
462     }
463 
464     return {std::move(direct_map_kernel), std::move(indirect_map_kernel)};
465   }
466 
467   struct LDSVariableReplacement {
468     GlobalVariable *SGV = nullptr;
469     DenseMap<GlobalVariable *, Constant *> LDSVarsToConstantGEP;
470   };
471 
472   // remap from lds global to a constantexpr gep to where it has been moved to
473   // for each kernel
474   // an array with an element for each kernel containing where the corresponding
475   // variable was remapped to
476 
477   static Constant *getAddressesOfVariablesInKernel(
478       LLVMContext &Ctx, ArrayRef<GlobalVariable *> Variables,
479       const DenseMap<GlobalVariable *, Constant *> &LDSVarsToConstantGEP) {
480     // Create a ConstantArray containing the address of each Variable within the
481     // kernel corresponding to LDSVarsToConstantGEP, or poison if that kernel
482     // does not allocate it
483     // TODO: Drop the ptrtoint conversion
484 
485     Type *I32 = Type::getInt32Ty(Ctx);
486 
487     ArrayType *KernelOffsetsType = ArrayType::get(I32, Variables.size());
488 
489     SmallVector<Constant *> Elements;
490     for (size_t i = 0; i < Variables.size(); i++) {
491       GlobalVariable *GV = Variables[i];
492       auto ConstantGepIt = LDSVarsToConstantGEP.find(GV);
493       if (ConstantGepIt != LDSVarsToConstantGEP.end()) {
494         auto elt = ConstantExpr::getPtrToInt(ConstantGepIt->second, I32);
495         Elements.push_back(elt);
496       } else {
497         Elements.push_back(PoisonValue::get(I32));
498       }
499     }
500     return ConstantArray::get(KernelOffsetsType, Elements);
501   }
502 
503   static GlobalVariable *buildLookupTable(
504       Module &M, ArrayRef<GlobalVariable *> Variables,
505       ArrayRef<Function *> kernels,
506       DenseMap<Function *, LDSVariableReplacement> &KernelToReplacement) {
507     if (Variables.empty()) {
508       return nullptr;
509     }
510     LLVMContext &Ctx = M.getContext();
511 
512     const size_t NumberVariables = Variables.size();
513     const size_t NumberKernels = kernels.size();
514 
515     ArrayType *KernelOffsetsType =
516         ArrayType::get(Type::getInt32Ty(Ctx), NumberVariables);
517 
518     ArrayType *AllKernelsOffsetsType =
519         ArrayType::get(KernelOffsetsType, NumberKernels);
520 
521     Constant *Missing = PoisonValue::get(KernelOffsetsType);
522     std::vector<Constant *> overallConstantExprElts(NumberKernels);
523     for (size_t i = 0; i < NumberKernels; i++) {
524       auto Replacement = KernelToReplacement.find(kernels[i]);
525       overallConstantExprElts[i] =
526           (Replacement == KernelToReplacement.end())
527               ? Missing
528               : getAddressesOfVariablesInKernel(
529                     Ctx, Variables, Replacement->second.LDSVarsToConstantGEP);
530     }
531 
532     Constant *init =
533         ConstantArray::get(AllKernelsOffsetsType, overallConstantExprElts);
534 
535     return new GlobalVariable(
536         M, AllKernelsOffsetsType, true, GlobalValue::InternalLinkage, init,
537         "llvm.amdgcn.lds.offset.table", nullptr, GlobalValue::NotThreadLocal,
538         AMDGPUAS::CONSTANT_ADDRESS);
539   }
540 
541   void replaceUseWithTableLookup(Module &M, IRBuilder<> &Builder,
542                                  GlobalVariable *LookupTable,
543                                  GlobalVariable *GV, Use &U,
544                                  Value *OptionalIndex) {
545     // Table is a constant array of the same length as OrderedKernels
546     LLVMContext &Ctx = M.getContext();
547     Type *I32 = Type::getInt32Ty(Ctx);
548     auto *I = cast<Instruction>(U.getUser());
549 
550     Value *tableKernelIndex = getTableLookupKernelIndex(M, I->getFunction());
551 
552     if (auto *Phi = dyn_cast<PHINode>(I)) {
553       BasicBlock *BB = Phi->getIncomingBlock(U);
554       Builder.SetInsertPoint(&(*(BB->getFirstInsertionPt())));
555     } else {
556       Builder.SetInsertPoint(I);
557     }
558 
559     SmallVector<Value *, 3> GEPIdx = {
560         ConstantInt::get(I32, 0),
561         tableKernelIndex,
562     };
563     if (OptionalIndex)
564       GEPIdx.push_back(OptionalIndex);
565 
566     Value *Address = Builder.CreateInBoundsGEP(
567         LookupTable->getValueType(), LookupTable, GEPIdx, GV->getName());
568 
569     Value *loaded = Builder.CreateLoad(I32, Address);
570 
571     Value *replacement =
572         Builder.CreateIntToPtr(loaded, GV->getType(), GV->getName());
573 
574     U.set(replacement);
575   }
576 
577   void replaceUsesInInstructionsWithTableLookup(
578       Module &M, ArrayRef<GlobalVariable *> ModuleScopeVariables,
579       GlobalVariable *LookupTable) {
580 
581     LLVMContext &Ctx = M.getContext();
582     IRBuilder<> Builder(Ctx);
583     Type *I32 = Type::getInt32Ty(Ctx);
584 
585     for (size_t Index = 0; Index < ModuleScopeVariables.size(); Index++) {
586       auto *GV = ModuleScopeVariables[Index];
587 
588       for (Use &U : make_early_inc_range(GV->uses())) {
589         auto *I = dyn_cast<Instruction>(U.getUser());
590         if (!I)
591           continue;
592 
593         replaceUseWithTableLookup(M, Builder, LookupTable, GV, U,
594                                   ConstantInt::get(I32, Index));
595       }
596     }
597   }
598 
599   static DenseSet<Function *> kernelsThatIndirectlyAccessAnyOfPassedVariables(
600       Module &M, LDSUsesInfoTy &LDSUsesInfo,
601       DenseSet<GlobalVariable *> const &VariableSet) {
602 
603     DenseSet<Function *> KernelSet;
604 
605     if (VariableSet.empty())
606       return KernelSet;
607 
608     for (Function &Func : M.functions()) {
609       if (Func.isDeclaration() || !isKernelLDS(&Func))
610         continue;
611       for (GlobalVariable *GV : LDSUsesInfo.indirect_access[&Func]) {
612         if (VariableSet.contains(GV)) {
613           KernelSet.insert(&Func);
614           break;
615         }
616       }
617     }
618 
619     return KernelSet;
620   }
621 
622   static GlobalVariable *
623   chooseBestVariableForModuleStrategy(const DataLayout &DL,
624                                       VariableFunctionMap &LDSVars) {
625     // Find the global variable with the most indirect uses from kernels
626 
627     struct CandidateTy {
628       GlobalVariable *GV = nullptr;
629       size_t UserCount = 0;
630       size_t Size = 0;
631 
632       CandidateTy() = default;
633 
634       CandidateTy(GlobalVariable *GV, uint64_t UserCount, uint64_t AllocSize)
635           : GV(GV), UserCount(UserCount), Size(AllocSize) {}
636 
637       bool operator<(const CandidateTy &Other) const {
638         // Fewer users makes module scope variable less attractive
639         if (UserCount < Other.UserCount) {
640           return true;
641         }
642         if (UserCount > Other.UserCount) {
643           return false;
644         }
645 
646         // Bigger makes module scope variable less attractive
647         if (Size < Other.Size) {
648           return false;
649         }
650 
651         if (Size > Other.Size) {
652           return true;
653         }
654 
655         // Arbitrary but consistent
656         return GV->getName() < Other.GV->getName();
657       }
658     };
659 
660     CandidateTy MostUsed;
661 
662     for (auto &K : LDSVars) {
663       GlobalVariable *GV = K.first;
664       if (K.second.size() <= 1) {
665         // A variable reachable by only one kernel is best lowered with kernel
666         // strategy
667         continue;
668       }
669       CandidateTy Candidate(
670           GV, K.second.size(),
671           DL.getTypeAllocSize(GV->getValueType()).getFixedValue());
672       if (MostUsed < Candidate)
673         MostUsed = Candidate;
674     }
675 
676     return MostUsed.GV;
677   }
678 
679   static void recordLDSAbsoluteAddress(Module *M, GlobalVariable *GV,
680                                        uint32_t Address) {
681     // Write the specified address into metadata where it can be retrieved by
682     // the assembler. Format is a half open range, [Address Address+1)
683     LLVMContext &Ctx = M->getContext();
684     auto *IntTy =
685         M->getDataLayout().getIntPtrType(Ctx, AMDGPUAS::LOCAL_ADDRESS);
686     auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntTy, Address));
687     auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntTy, Address + 1));
688     GV->setMetadata(LLVMContext::MD_absolute_symbol,
689                     MDNode::get(Ctx, {MinC, MaxC}));
690   }
691 
692   DenseMap<Function *, Value *> tableKernelIndexCache;
693   Value *getTableLookupKernelIndex(Module &M, Function *F) {
694     // Accesses from a function use the amdgcn_lds_kernel_id intrinsic which
695     // lowers to a read from a live in register. Emit it once in the entry
696     // block to spare deduplicating it later.
697     auto [It, Inserted] = tableKernelIndexCache.try_emplace(F);
698     if (Inserted) {
699       Function *Decl =
700           Intrinsic::getDeclaration(&M, Intrinsic::amdgcn_lds_kernel_id, {});
701 
702       auto InsertAt = F->getEntryBlock().getFirstNonPHIOrDbgOrAlloca();
703       IRBuilder<> Builder(&*InsertAt);
704 
705       It->second = Builder.CreateCall(Decl, {});
706     }
707 
708     return It->second;
709   }
710 
711   static std::vector<Function *> assignLDSKernelIDToEachKernel(
712       Module *M, DenseSet<Function *> const &KernelsThatAllocateTableLDS,
713       DenseSet<Function *> const &KernelsThatIndirectlyAllocateDynamicLDS) {
714     // Associate kernels in the set with an arbirary but reproducible order and
715     // annotate them with that order in metadata. This metadata is recognised by
716     // the backend and lowered to a SGPR which can be read from using
717     // amdgcn_lds_kernel_id.
718 
719     std::vector<Function *> OrderedKernels;
720     if (!KernelsThatAllocateTableLDS.empty() ||
721         !KernelsThatIndirectlyAllocateDynamicLDS.empty()) {
722 
723       for (Function &Func : M->functions()) {
724         if (Func.isDeclaration())
725           continue;
726         if (!isKernelLDS(&Func))
727           continue;
728 
729         if (KernelsThatAllocateTableLDS.contains(&Func) ||
730             KernelsThatIndirectlyAllocateDynamicLDS.contains(&Func)) {
731           assert(Func.hasName()); // else fatal error earlier
732           OrderedKernels.push_back(&Func);
733         }
734       }
735 
736       // Put them in an arbitrary but reproducible order
737       OrderedKernels = sortByName(std::move(OrderedKernels));
738 
739       // Annotate the kernels with their order in this vector
740       LLVMContext &Ctx = M->getContext();
741       IRBuilder<> Builder(Ctx);
742 
743       if (OrderedKernels.size() > UINT32_MAX) {
744         // 32 bit keeps it in one SGPR. > 2**32 kernels won't fit on the GPU
745         report_fatal_error("Unimplemented LDS lowering for > 2**32 kernels");
746       }
747 
748       for (size_t i = 0; i < OrderedKernels.size(); i++) {
749         Metadata *AttrMDArgs[1] = {
750             ConstantAsMetadata::get(Builder.getInt32(i)),
751         };
752         OrderedKernels[i]->setMetadata("llvm.amdgcn.lds.kernel.id",
753                                        MDNode::get(Ctx, AttrMDArgs));
754       }
755     }
756     return OrderedKernels;
757   }
758 
759   static void partitionVariablesIntoIndirectStrategies(
760       Module &M, LDSUsesInfoTy const &LDSUsesInfo,
761       VariableFunctionMap &LDSToKernelsThatNeedToAccessItIndirectly,
762       DenseSet<GlobalVariable *> &ModuleScopeVariables,
763       DenseSet<GlobalVariable *> &TableLookupVariables,
764       DenseSet<GlobalVariable *> &KernelAccessVariables,
765       DenseSet<GlobalVariable *> &DynamicVariables) {
766 
767     GlobalVariable *HybridModuleRoot =
768         LoweringKindLoc != LoweringKind::hybrid
769             ? nullptr
770             : chooseBestVariableForModuleStrategy(
771                   M.getDataLayout(), LDSToKernelsThatNeedToAccessItIndirectly);
772 
773     DenseSet<Function *> const EmptySet;
774     DenseSet<Function *> const &HybridModuleRootKernels =
775         HybridModuleRoot
776             ? LDSToKernelsThatNeedToAccessItIndirectly[HybridModuleRoot]
777             : EmptySet;
778 
779     for (auto &K : LDSToKernelsThatNeedToAccessItIndirectly) {
780       // Each iteration of this loop assigns exactly one global variable to
781       // exactly one of the implementation strategies.
782 
783       GlobalVariable *GV = K.first;
784       assert(AMDGPU::isLDSVariableToLower(*GV));
785       assert(K.second.size() != 0);
786 
787       if (AMDGPU::isDynamicLDS(*GV)) {
788         DynamicVariables.insert(GV);
789         continue;
790       }
791 
792       switch (LoweringKindLoc) {
793       case LoweringKind::module:
794         ModuleScopeVariables.insert(GV);
795         break;
796 
797       case LoweringKind::table:
798         TableLookupVariables.insert(GV);
799         break;
800 
801       case LoweringKind::kernel:
802         if (K.second.size() == 1) {
803           KernelAccessVariables.insert(GV);
804         } else {
805           report_fatal_error(
806               "cannot lower LDS '" + GV->getName() +
807               "' to kernel access as it is reachable from multiple kernels");
808         }
809         break;
810 
811       case LoweringKind::hybrid: {
812         if (GV == HybridModuleRoot) {
813           assert(K.second.size() != 1);
814           ModuleScopeVariables.insert(GV);
815         } else if (K.second.size() == 1) {
816           KernelAccessVariables.insert(GV);
817         } else if (set_is_subset(K.second, HybridModuleRootKernels)) {
818           ModuleScopeVariables.insert(GV);
819         } else {
820           TableLookupVariables.insert(GV);
821         }
822         break;
823       }
824       }
825     }
826 
827     // All LDS variables accessed indirectly have now been partitioned into
828     // the distinct lowering strategies.
829     assert(ModuleScopeVariables.size() + TableLookupVariables.size() +
830                KernelAccessVariables.size() + DynamicVariables.size() ==
831            LDSToKernelsThatNeedToAccessItIndirectly.size());
832   }
833 
834   static GlobalVariable *lowerModuleScopeStructVariables(
835       Module &M, DenseSet<GlobalVariable *> const &ModuleScopeVariables,
836       DenseSet<Function *> const &KernelsThatAllocateModuleLDS) {
837     // Create a struct to hold the ModuleScopeVariables
838     // Replace all uses of those variables from non-kernel functions with the
839     // new struct instance Replace only the uses from kernel functions that will
840     // allocate this instance. That is a space optimisation - kernels that use a
841     // subset of the module scope struct and do not need to allocate it for
842     // indirect calls will only allocate the subset they use (they do so as part
843     // of the per-kernel lowering).
844     if (ModuleScopeVariables.empty()) {
845       return nullptr;
846     }
847 
848     LLVMContext &Ctx = M.getContext();
849 
850     LDSVariableReplacement ModuleScopeReplacement =
851         createLDSVariableReplacement(M, "llvm.amdgcn.module.lds",
852                                      ModuleScopeVariables);
853 
854     appendToCompilerUsed(M, {static_cast<GlobalValue *>(
855                                 ConstantExpr::getPointerBitCastOrAddrSpaceCast(
856                                     cast<Constant>(ModuleScopeReplacement.SGV),
857                                     PointerType::getUnqual(Ctx)))});
858 
859     // module.lds will be allocated at zero in any kernel that allocates it
860     recordLDSAbsoluteAddress(&M, ModuleScopeReplacement.SGV, 0);
861 
862     // historic
863     removeLocalVarsFromUsedLists(M, ModuleScopeVariables);
864 
865     // Replace all uses of module scope variable from non-kernel functions
866     replaceLDSVariablesWithStruct(
867         M, ModuleScopeVariables, ModuleScopeReplacement, [&](Use &U) {
868           Instruction *I = dyn_cast<Instruction>(U.getUser());
869           if (!I) {
870             return false;
871           }
872           Function *F = I->getFunction();
873           return !isKernelLDS(F);
874         });
875 
876     // Replace uses of module scope variable from kernel functions that
877     // allocate the module scope variable, otherwise leave them unchanged
878     // Record on each kernel whether the module scope global is used by it
879 
880     for (Function &Func : M.functions()) {
881       if (Func.isDeclaration() || !isKernelLDS(&Func))
882         continue;
883 
884       if (KernelsThatAllocateModuleLDS.contains(&Func)) {
885         replaceLDSVariablesWithStruct(
886             M, ModuleScopeVariables, ModuleScopeReplacement, [&](Use &U) {
887               Instruction *I = dyn_cast<Instruction>(U.getUser());
888               if (!I) {
889                 return false;
890               }
891               Function *F = I->getFunction();
892               return F == &Func;
893             });
894 
895         markUsedByKernel(&Func, ModuleScopeReplacement.SGV);
896       }
897     }
898 
899     return ModuleScopeReplacement.SGV;
900   }
901 
902   static DenseMap<Function *, LDSVariableReplacement>
903   lowerKernelScopeStructVariables(
904       Module &M, LDSUsesInfoTy &LDSUsesInfo,
905       DenseSet<GlobalVariable *> const &ModuleScopeVariables,
906       DenseSet<Function *> const &KernelsThatAllocateModuleLDS,
907       GlobalVariable *MaybeModuleScopeStruct) {
908 
909     // Create a struct for each kernel for the non-module-scope variables.
910 
911     DenseMap<Function *, LDSVariableReplacement> KernelToReplacement;
912     for (Function &Func : M.functions()) {
913       if (Func.isDeclaration() || !isKernelLDS(&Func))
914         continue;
915 
916       DenseSet<GlobalVariable *> KernelUsedVariables;
917       // Allocating variables that are used directly in this struct to get
918       // alignment aware allocation and predictable frame size.
919       for (auto &v : LDSUsesInfo.direct_access[&Func]) {
920         if (!AMDGPU::isDynamicLDS(*v)) {
921           KernelUsedVariables.insert(v);
922         }
923       }
924 
925       // Allocating variables that are accessed indirectly so that a lookup of
926       // this struct instance can find them from nested functions.
927       for (auto &v : LDSUsesInfo.indirect_access[&Func]) {
928         if (!AMDGPU::isDynamicLDS(*v)) {
929           KernelUsedVariables.insert(v);
930         }
931       }
932 
933       // Variables allocated in module lds must all resolve to that struct,
934       // not to the per-kernel instance.
935       if (KernelsThatAllocateModuleLDS.contains(&Func)) {
936         for (GlobalVariable *v : ModuleScopeVariables) {
937           KernelUsedVariables.erase(v);
938         }
939       }
940 
941       if (KernelUsedVariables.empty()) {
942         // Either used no LDS, or the LDS it used was all in the module struct
943         // or dynamically sized
944         continue;
945       }
946 
947       // The association between kernel function and LDS struct is done by
948       // symbol name, which only works if the function in question has a
949       // name This is not expected to be a problem in practice as kernels
950       // are called by name making anonymous ones (which are named by the
951       // backend) difficult to use. This does mean that llvm test cases need
952       // to name the kernels.
953       if (!Func.hasName()) {
954         report_fatal_error("Anonymous kernels cannot use LDS variables");
955       }
956 
957       std::string VarName =
958           (Twine("llvm.amdgcn.kernel.") + Func.getName() + ".lds").str();
959 
960       auto Replacement =
961           createLDSVariableReplacement(M, VarName, KernelUsedVariables);
962 
963       // If any indirect uses, create a direct use to ensure allocation
964       // TODO: Simpler to unconditionally mark used but that regresses
965       // codegen in test/CodeGen/AMDGPU/noclobber-barrier.ll
966       auto Accesses = LDSUsesInfo.indirect_access.find(&Func);
967       if ((Accesses != LDSUsesInfo.indirect_access.end()) &&
968           !Accesses->second.empty())
969         markUsedByKernel(&Func, Replacement.SGV);
970 
971       // remove preserves existing codegen
972       removeLocalVarsFromUsedLists(M, KernelUsedVariables);
973       KernelToReplacement[&Func] = Replacement;
974 
975       // Rewrite uses within kernel to the new struct
976       replaceLDSVariablesWithStruct(
977           M, KernelUsedVariables, Replacement, [&Func](Use &U) {
978             Instruction *I = dyn_cast<Instruction>(U.getUser());
979             return I && I->getFunction() == &Func;
980           });
981     }
982     return KernelToReplacement;
983   }
984 
985   static GlobalVariable *
986   buildRepresentativeDynamicLDSInstance(Module &M, LDSUsesInfoTy &LDSUsesInfo,
987                                         Function *func) {
988     // Create a dynamic lds variable with a name associated with the passed
989     // function that has the maximum alignment of any dynamic lds variable
990     // reachable from this kernel. Dynamic LDS is allocated after the static LDS
991     // allocation, possibly after alignment padding. The representative variable
992     // created here has the maximum alignment of any other dynamic variable
993     // reachable by that kernel. All dynamic LDS variables are allocated at the
994     // same address in each kernel in order to provide the documented aliasing
995     // semantics. Setting the alignment here allows this IR pass to accurately
996     // predict the exact constant at which it will be allocated.
997 
998     assert(isKernelLDS(func));
999 
1000     LLVMContext &Ctx = M.getContext();
1001     const DataLayout &DL = M.getDataLayout();
1002     Align MaxDynamicAlignment(1);
1003 
1004     auto UpdateMaxAlignment = [&MaxDynamicAlignment, &DL](GlobalVariable *GV) {
1005       if (AMDGPU::isDynamicLDS(*GV)) {
1006         MaxDynamicAlignment =
1007             std::max(MaxDynamicAlignment, AMDGPU::getAlign(DL, GV));
1008       }
1009     };
1010 
1011     for (GlobalVariable *GV : LDSUsesInfo.indirect_access[func]) {
1012       UpdateMaxAlignment(GV);
1013     }
1014 
1015     for (GlobalVariable *GV : LDSUsesInfo.direct_access[func]) {
1016       UpdateMaxAlignment(GV);
1017     }
1018 
1019     assert(func->hasName()); // Checked by caller
1020     auto emptyCharArray = ArrayType::get(Type::getInt8Ty(Ctx), 0);
1021     GlobalVariable *N = new GlobalVariable(
1022         M, emptyCharArray, false, GlobalValue::ExternalLinkage, nullptr,
1023         Twine("llvm.amdgcn." + func->getName() + ".dynlds"), nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS,
1024         false);
1025     N->setAlignment(MaxDynamicAlignment);
1026 
1027     assert(AMDGPU::isDynamicLDS(*N));
1028     return N;
1029   }
1030 
1031   DenseMap<Function *, GlobalVariable *> lowerDynamicLDSVariables(
1032       Module &M, LDSUsesInfoTy &LDSUsesInfo,
1033       DenseSet<Function *> const &KernelsThatIndirectlyAllocateDynamicLDS,
1034       DenseSet<GlobalVariable *> const &DynamicVariables,
1035       std::vector<Function *> const &OrderedKernels) {
1036     DenseMap<Function *, GlobalVariable *> KernelToCreatedDynamicLDS;
1037     if (!KernelsThatIndirectlyAllocateDynamicLDS.empty()) {
1038       LLVMContext &Ctx = M.getContext();
1039       IRBuilder<> Builder(Ctx);
1040       Type *I32 = Type::getInt32Ty(Ctx);
1041 
1042       std::vector<Constant *> newDynamicLDS;
1043 
1044       // Table is built in the same order as OrderedKernels
1045       for (auto &func : OrderedKernels) {
1046 
1047         if (KernelsThatIndirectlyAllocateDynamicLDS.contains(func)) {
1048           assert(isKernelLDS(func));
1049           if (!func->hasName()) {
1050             report_fatal_error("Anonymous kernels cannot use LDS variables");
1051           }
1052 
1053           GlobalVariable *N =
1054               buildRepresentativeDynamicLDSInstance(M, LDSUsesInfo, func);
1055 
1056           KernelToCreatedDynamicLDS[func] = N;
1057 
1058           markUsedByKernel(func, N);
1059 
1060           auto emptyCharArray = ArrayType::get(Type::getInt8Ty(Ctx), 0);
1061           auto GEP = ConstantExpr::getGetElementPtr(
1062               emptyCharArray, N, ConstantInt::get(I32, 0), true);
1063           newDynamicLDS.push_back(ConstantExpr::getPtrToInt(GEP, I32));
1064         } else {
1065           newDynamicLDS.push_back(PoisonValue::get(I32));
1066         }
1067       }
1068       assert(OrderedKernels.size() == newDynamicLDS.size());
1069 
1070       ArrayType *t = ArrayType::get(I32, newDynamicLDS.size());
1071       Constant *init = ConstantArray::get(t, newDynamicLDS);
1072       GlobalVariable *table = new GlobalVariable(
1073           M, t, true, GlobalValue::InternalLinkage, init,
1074           "llvm.amdgcn.dynlds.offset.table", nullptr,
1075           GlobalValue::NotThreadLocal, AMDGPUAS::CONSTANT_ADDRESS);
1076 
1077       for (GlobalVariable *GV : DynamicVariables) {
1078         for (Use &U : make_early_inc_range(GV->uses())) {
1079           auto *I = dyn_cast<Instruction>(U.getUser());
1080           if (!I)
1081             continue;
1082           if (isKernelLDS(I->getFunction()))
1083             continue;
1084 
1085           replaceUseWithTableLookup(M, Builder, table, GV, U, nullptr);
1086         }
1087       }
1088     }
1089     return KernelToCreatedDynamicLDS;
1090   }
1091 
1092   bool runOnModule(Module &M) {
1093     CallGraph CG = CallGraph(M);
1094     bool Changed = superAlignLDSGlobals(M);
1095 
1096     Changed |= eliminateConstantExprUsesOfLDSFromAllInstructions(M);
1097 
1098     Changed = true; // todo: narrow this down
1099 
1100     // For each kernel, what variables does it access directly or through
1101     // callees
1102     LDSUsesInfoTy LDSUsesInfo = getTransitiveUsesOfLDS(CG, M);
1103 
1104     // For each variable accessed through callees, which kernels access it
1105     VariableFunctionMap LDSToKernelsThatNeedToAccessItIndirectly;
1106     for (auto &K : LDSUsesInfo.indirect_access) {
1107       Function *F = K.first;
1108       assert(isKernelLDS(F));
1109       for (GlobalVariable *GV : K.second) {
1110         LDSToKernelsThatNeedToAccessItIndirectly[GV].insert(F);
1111       }
1112     }
1113 
1114     // Partition variables accessed indirectly into the different strategies
1115     DenseSet<GlobalVariable *> ModuleScopeVariables;
1116     DenseSet<GlobalVariable *> TableLookupVariables;
1117     DenseSet<GlobalVariable *> KernelAccessVariables;
1118     DenseSet<GlobalVariable *> DynamicVariables;
1119     partitionVariablesIntoIndirectStrategies(
1120         M, LDSUsesInfo, LDSToKernelsThatNeedToAccessItIndirectly,
1121         ModuleScopeVariables, TableLookupVariables, KernelAccessVariables,
1122         DynamicVariables);
1123 
1124     // If the kernel accesses a variable that is going to be stored in the
1125     // module instance through a call then that kernel needs to allocate the
1126     // module instance
1127     const DenseSet<Function *> KernelsThatAllocateModuleLDS =
1128         kernelsThatIndirectlyAccessAnyOfPassedVariables(M, LDSUsesInfo,
1129                                                         ModuleScopeVariables);
1130     const DenseSet<Function *> KernelsThatAllocateTableLDS =
1131         kernelsThatIndirectlyAccessAnyOfPassedVariables(M, LDSUsesInfo,
1132                                                         TableLookupVariables);
1133 
1134     const DenseSet<Function *> KernelsThatIndirectlyAllocateDynamicLDS =
1135         kernelsThatIndirectlyAccessAnyOfPassedVariables(M, LDSUsesInfo,
1136                                                         DynamicVariables);
1137 
1138     GlobalVariable *MaybeModuleScopeStruct = lowerModuleScopeStructVariables(
1139         M, ModuleScopeVariables, KernelsThatAllocateModuleLDS);
1140 
1141     DenseMap<Function *, LDSVariableReplacement> KernelToReplacement =
1142         lowerKernelScopeStructVariables(M, LDSUsesInfo, ModuleScopeVariables,
1143                                         KernelsThatAllocateModuleLDS,
1144                                         MaybeModuleScopeStruct);
1145 
1146     // Lower zero cost accesses to the kernel instances just created
1147     for (auto &GV : KernelAccessVariables) {
1148       auto &funcs = LDSToKernelsThatNeedToAccessItIndirectly[GV];
1149       assert(funcs.size() == 1); // Only one kernel can access it
1150       LDSVariableReplacement Replacement =
1151           KernelToReplacement[*(funcs.begin())];
1152 
1153       DenseSet<GlobalVariable *> Vec;
1154       Vec.insert(GV);
1155 
1156       replaceLDSVariablesWithStruct(M, Vec, Replacement, [](Use &U) {
1157         return isa<Instruction>(U.getUser());
1158       });
1159     }
1160 
1161     // The ith element of this vector is kernel id i
1162     std::vector<Function *> OrderedKernels =
1163         assignLDSKernelIDToEachKernel(&M, KernelsThatAllocateTableLDS,
1164                                       KernelsThatIndirectlyAllocateDynamicLDS);
1165 
1166     if (!KernelsThatAllocateTableLDS.empty()) {
1167       LLVMContext &Ctx = M.getContext();
1168       IRBuilder<> Builder(Ctx);
1169 
1170       // The order must be consistent between lookup table and accesses to
1171       // lookup table
1172       auto TableLookupVariablesOrdered =
1173           sortByName(std::vector<GlobalVariable *>(TableLookupVariables.begin(),
1174                                                    TableLookupVariables.end()));
1175 
1176       GlobalVariable *LookupTable = buildLookupTable(
1177           M, TableLookupVariablesOrdered, OrderedKernels, KernelToReplacement);
1178       replaceUsesInInstructionsWithTableLookup(M, TableLookupVariablesOrdered,
1179                                                LookupTable);
1180     }
1181 
1182     DenseMap<Function *, GlobalVariable *> KernelToCreatedDynamicLDS =
1183         lowerDynamicLDSVariables(M, LDSUsesInfo,
1184                                  KernelsThatIndirectlyAllocateDynamicLDS,
1185                                  DynamicVariables, OrderedKernels);
1186 
1187     // All kernel frames have been allocated. Calculate and record the
1188     // addresses.
1189     {
1190       const DataLayout &DL = M.getDataLayout();
1191 
1192       for (Function &Func : M.functions()) {
1193         if (Func.isDeclaration() || !isKernelLDS(&Func))
1194           continue;
1195 
1196         // All three of these are optional. The first variable is allocated at
1197         // zero. They are allocated by AMDGPUMachineFunction as one block.
1198         // Layout:
1199         //{
1200         //  module.lds
1201         //  alignment padding
1202         //  kernel instance
1203         //  alignment padding
1204         //  dynamic lds variables
1205         //}
1206 
1207         const bool AllocateModuleScopeStruct =
1208             MaybeModuleScopeStruct &&
1209             KernelsThatAllocateModuleLDS.contains(&Func);
1210 
1211         auto Replacement = KernelToReplacement.find(&Func);
1212         const bool AllocateKernelScopeStruct =
1213             Replacement != KernelToReplacement.end();
1214 
1215         const bool AllocateDynamicVariable =
1216             KernelToCreatedDynamicLDS.contains(&Func);
1217 
1218         uint32_t Offset = 0;
1219 
1220         if (AllocateModuleScopeStruct) {
1221           // Allocated at zero, recorded once on construction, not once per
1222           // kernel
1223           Offset += DL.getTypeAllocSize(MaybeModuleScopeStruct->getValueType());
1224         }
1225 
1226         if (AllocateKernelScopeStruct) {
1227           GlobalVariable *KernelStruct = Replacement->second.SGV;
1228           Offset = alignTo(Offset, AMDGPU::getAlign(DL, KernelStruct));
1229           recordLDSAbsoluteAddress(&M, KernelStruct, Offset);
1230           Offset += DL.getTypeAllocSize(KernelStruct->getValueType());
1231         }
1232 
1233         // If there is dynamic allocation, the alignment needed is included in
1234         // the static frame size. There may be no reference to the dynamic
1235         // variable in the kernel itself, so without including it here, that
1236         // alignment padding could be missed.
1237         if (AllocateDynamicVariable) {
1238           GlobalVariable *DynamicVariable = KernelToCreatedDynamicLDS[&Func];
1239           Offset = alignTo(Offset, AMDGPU::getAlign(DL, DynamicVariable));
1240           recordLDSAbsoluteAddress(&M, DynamicVariable, Offset);
1241         }
1242 
1243         if (Offset != 0) {
1244           (void)TM; // TODO: Account for target maximum LDS
1245           std::string Buffer;
1246           raw_string_ostream SS{Buffer};
1247           SS << format("%u", Offset);
1248 
1249           // Instead of explictly marking kernels that access dynamic variables
1250           // using special case metadata, annotate with min-lds == max-lds, i.e.
1251           // that there is no more space available for allocating more static
1252           // LDS variables. That is the right condition to prevent allocating
1253           // more variables which would collide with the addresses assigned to
1254           // dynamic variables.
1255           if (AllocateDynamicVariable)
1256             SS << format(",%u", Offset);
1257 
1258           Func.addFnAttr("amdgpu-lds-size", Buffer);
1259         }
1260       }
1261     }
1262 
1263     for (auto &GV : make_early_inc_range(M.globals()))
1264       if (AMDGPU::isLDSVariableToLower(GV)) {
1265         // probably want to remove from used lists
1266         GV.removeDeadConstantUsers();
1267         if (GV.use_empty())
1268           GV.eraseFromParent();
1269       }
1270 
1271     return Changed;
1272   }
1273 
1274 private:
1275   // Increase the alignment of LDS globals if necessary to maximise the chance
1276   // that we can use aligned LDS instructions to access them.
1277   static bool superAlignLDSGlobals(Module &M) {
1278     const DataLayout &DL = M.getDataLayout();
1279     bool Changed = false;
1280     if (!SuperAlignLDSGlobals) {
1281       return Changed;
1282     }
1283 
1284     for (auto &GV : M.globals()) {
1285       if (GV.getType()->getPointerAddressSpace() != AMDGPUAS::LOCAL_ADDRESS) {
1286         // Only changing alignment of LDS variables
1287         continue;
1288       }
1289       if (!GV.hasInitializer()) {
1290         // cuda/hip extern __shared__ variable, leave alignment alone
1291         continue;
1292       }
1293 
1294       Align Alignment = AMDGPU::getAlign(DL, &GV);
1295       TypeSize GVSize = DL.getTypeAllocSize(GV.getValueType());
1296 
1297       if (GVSize > 8) {
1298         // We might want to use a b96 or b128 load/store
1299         Alignment = std::max(Alignment, Align(16));
1300       } else if (GVSize > 4) {
1301         // We might want to use a b64 load/store
1302         Alignment = std::max(Alignment, Align(8));
1303       } else if (GVSize > 2) {
1304         // We might want to use a b32 load/store
1305         Alignment = std::max(Alignment, Align(4));
1306       } else if (GVSize > 1) {
1307         // We might want to use a b16 load/store
1308         Alignment = std::max(Alignment, Align(2));
1309       }
1310 
1311       if (Alignment != AMDGPU::getAlign(DL, &GV)) {
1312         Changed = true;
1313         GV.setAlignment(Alignment);
1314       }
1315     }
1316     return Changed;
1317   }
1318 
1319   static LDSVariableReplacement createLDSVariableReplacement(
1320       Module &M, std::string VarName,
1321       DenseSet<GlobalVariable *> const &LDSVarsToTransform) {
1322     // Create a struct instance containing LDSVarsToTransform and map from those
1323     // variables to ConstantExprGEP
1324     // Variables may be introduced to meet alignment requirements. No aliasing
1325     // metadata is useful for these as they have no uses. Erased before return.
1326 
1327     LLVMContext &Ctx = M.getContext();
1328     const DataLayout &DL = M.getDataLayout();
1329     assert(!LDSVarsToTransform.empty());
1330 
1331     SmallVector<OptimizedStructLayoutField, 8> LayoutFields;
1332     LayoutFields.reserve(LDSVarsToTransform.size());
1333     {
1334       // The order of fields in this struct depends on the order of
1335       // varables in the argument which varies when changing how they
1336       // are identified, leading to spurious test breakage.
1337       auto Sorted = sortByName(std::vector<GlobalVariable *>(
1338           LDSVarsToTransform.begin(), LDSVarsToTransform.end()));
1339 
1340       for (GlobalVariable *GV : Sorted) {
1341         OptimizedStructLayoutField F(GV,
1342                                      DL.getTypeAllocSize(GV->getValueType()),
1343                                      AMDGPU::getAlign(DL, GV));
1344         LayoutFields.emplace_back(F);
1345       }
1346     }
1347 
1348     performOptimizedStructLayout(LayoutFields);
1349 
1350     std::vector<GlobalVariable *> LocalVars;
1351     BitVector IsPaddingField;
1352     LocalVars.reserve(LDSVarsToTransform.size()); // will be at least this large
1353     IsPaddingField.reserve(LDSVarsToTransform.size());
1354     {
1355       uint64_t CurrentOffset = 0;
1356       for (size_t I = 0; I < LayoutFields.size(); I++) {
1357         GlobalVariable *FGV = static_cast<GlobalVariable *>(
1358             const_cast<void *>(LayoutFields[I].Id));
1359         Align DataAlign = LayoutFields[I].Alignment;
1360 
1361         uint64_t DataAlignV = DataAlign.value();
1362         if (uint64_t Rem = CurrentOffset % DataAlignV) {
1363           uint64_t Padding = DataAlignV - Rem;
1364 
1365           // Append an array of padding bytes to meet alignment requested
1366           // Note (o +      (a - (o % a)) ) % a == 0
1367           //      (offset + Padding       ) % align == 0
1368 
1369           Type *ATy = ArrayType::get(Type::getInt8Ty(Ctx), Padding);
1370           LocalVars.push_back(new GlobalVariable(
1371               M, ATy, false, GlobalValue::InternalLinkage,
1372               PoisonValue::get(ATy), "", nullptr, GlobalValue::NotThreadLocal,
1373               AMDGPUAS::LOCAL_ADDRESS, false));
1374           IsPaddingField.push_back(true);
1375           CurrentOffset += Padding;
1376         }
1377 
1378         LocalVars.push_back(FGV);
1379         IsPaddingField.push_back(false);
1380         CurrentOffset += LayoutFields[I].Size;
1381       }
1382     }
1383 
1384     std::vector<Type *> LocalVarTypes;
1385     LocalVarTypes.reserve(LocalVars.size());
1386     std::transform(
1387         LocalVars.cbegin(), LocalVars.cend(), std::back_inserter(LocalVarTypes),
1388         [](const GlobalVariable *V) -> Type * { return V->getValueType(); });
1389 
1390     StructType *LDSTy = StructType::create(Ctx, LocalVarTypes, VarName + ".t");
1391 
1392     Align StructAlign = AMDGPU::getAlign(DL, LocalVars[0]);
1393 
1394     GlobalVariable *SGV = new GlobalVariable(
1395         M, LDSTy, false, GlobalValue::InternalLinkage, PoisonValue::get(LDSTy),
1396         VarName, nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS,
1397         false);
1398     SGV->setAlignment(StructAlign);
1399 
1400     DenseMap<GlobalVariable *, Constant *> Map;
1401     Type *I32 = Type::getInt32Ty(Ctx);
1402     for (size_t I = 0; I < LocalVars.size(); I++) {
1403       GlobalVariable *GV = LocalVars[I];
1404       Constant *GEPIdx[] = {ConstantInt::get(I32, 0), ConstantInt::get(I32, I)};
1405       Constant *GEP = ConstantExpr::getGetElementPtr(LDSTy, SGV, GEPIdx, true);
1406       if (IsPaddingField[I]) {
1407         assert(GV->use_empty());
1408         GV->eraseFromParent();
1409       } else {
1410         Map[GV] = GEP;
1411       }
1412     }
1413     assert(Map.size() == LDSVarsToTransform.size());
1414     return {SGV, std::move(Map)};
1415   }
1416 
1417   template <typename PredicateTy>
1418   static void replaceLDSVariablesWithStruct(
1419       Module &M, DenseSet<GlobalVariable *> const &LDSVarsToTransformArg,
1420       const LDSVariableReplacement &Replacement, PredicateTy Predicate) {
1421     LLVMContext &Ctx = M.getContext();
1422     const DataLayout &DL = M.getDataLayout();
1423 
1424     // A hack... we need to insert the aliasing info in a predictable order for
1425     // lit tests. Would like to have them in a stable order already, ideally the
1426     // same order they get allocated, which might mean an ordered set container
1427     auto LDSVarsToTransform = sortByName(std::vector<GlobalVariable *>(
1428         LDSVarsToTransformArg.begin(), LDSVarsToTransformArg.end()));
1429 
1430     // Create alias.scope and their lists. Each field in the new structure
1431     // does not alias with all other fields.
1432     SmallVector<MDNode *> AliasScopes;
1433     SmallVector<Metadata *> NoAliasList;
1434     const size_t NumberVars = LDSVarsToTransform.size();
1435     if (NumberVars > 1) {
1436       MDBuilder MDB(Ctx);
1437       AliasScopes.reserve(NumberVars);
1438       MDNode *Domain = MDB.createAnonymousAliasScopeDomain();
1439       for (size_t I = 0; I < NumberVars; I++) {
1440         MDNode *Scope = MDB.createAnonymousAliasScope(Domain);
1441         AliasScopes.push_back(Scope);
1442       }
1443       NoAliasList.append(&AliasScopes[1], AliasScopes.end());
1444     }
1445 
1446     // Replace uses of ith variable with a constantexpr to the corresponding
1447     // field of the instance that will be allocated by AMDGPUMachineFunction
1448     for (size_t I = 0; I < NumberVars; I++) {
1449       GlobalVariable *GV = LDSVarsToTransform[I];
1450       Constant *GEP = Replacement.LDSVarsToConstantGEP.at(GV);
1451 
1452       GV->replaceUsesWithIf(GEP, Predicate);
1453 
1454       APInt APOff(DL.getIndexTypeSizeInBits(GEP->getType()), 0);
1455       GEP->stripAndAccumulateInBoundsConstantOffsets(DL, APOff);
1456       uint64_t Offset = APOff.getZExtValue();
1457 
1458       Align A =
1459           commonAlignment(Replacement.SGV->getAlign().valueOrOne(), Offset);
1460 
1461       if (I)
1462         NoAliasList[I - 1] = AliasScopes[I - 1];
1463       MDNode *NoAlias =
1464           NoAliasList.empty() ? nullptr : MDNode::get(Ctx, NoAliasList);
1465       MDNode *AliasScope =
1466           AliasScopes.empty() ? nullptr : MDNode::get(Ctx, {AliasScopes[I]});
1467 
1468       refineUsesAlignmentAndAA(GEP, A, DL, AliasScope, NoAlias);
1469     }
1470   }
1471 
1472   static void refineUsesAlignmentAndAA(Value *Ptr, Align A,
1473                                        const DataLayout &DL, MDNode *AliasScope,
1474                                        MDNode *NoAlias, unsigned MaxDepth = 5) {
1475     if (!MaxDepth || (A == 1 && !AliasScope))
1476       return;
1477 
1478     for (User *U : Ptr->users()) {
1479       if (auto *I = dyn_cast<Instruction>(U)) {
1480         if (AliasScope && I->mayReadOrWriteMemory()) {
1481           MDNode *AS = I->getMetadata(LLVMContext::MD_alias_scope);
1482           AS = (AS ? MDNode::getMostGenericAliasScope(AS, AliasScope)
1483                    : AliasScope);
1484           I->setMetadata(LLVMContext::MD_alias_scope, AS);
1485 
1486           MDNode *NA = I->getMetadata(LLVMContext::MD_noalias);
1487           NA = (NA ? MDNode::intersect(NA, NoAlias) : NoAlias);
1488           I->setMetadata(LLVMContext::MD_noalias, NA);
1489         }
1490       }
1491 
1492       if (auto *LI = dyn_cast<LoadInst>(U)) {
1493         LI->setAlignment(std::max(A, LI->getAlign()));
1494         continue;
1495       }
1496       if (auto *SI = dyn_cast<StoreInst>(U)) {
1497         if (SI->getPointerOperand() == Ptr)
1498           SI->setAlignment(std::max(A, SI->getAlign()));
1499         continue;
1500       }
1501       if (auto *AI = dyn_cast<AtomicRMWInst>(U)) {
1502         // None of atomicrmw operations can work on pointers, but let's
1503         // check it anyway in case it will or we will process ConstantExpr.
1504         if (AI->getPointerOperand() == Ptr)
1505           AI->setAlignment(std::max(A, AI->getAlign()));
1506         continue;
1507       }
1508       if (auto *AI = dyn_cast<AtomicCmpXchgInst>(U)) {
1509         if (AI->getPointerOperand() == Ptr)
1510           AI->setAlignment(std::max(A, AI->getAlign()));
1511         continue;
1512       }
1513       if (auto *GEP = dyn_cast<GetElementPtrInst>(U)) {
1514         unsigned BitWidth = DL.getIndexTypeSizeInBits(GEP->getType());
1515         APInt Off(BitWidth, 0);
1516         if (GEP->getPointerOperand() == Ptr) {
1517           Align GA;
1518           if (GEP->accumulateConstantOffset(DL, Off))
1519             GA = commonAlignment(A, Off.getLimitedValue());
1520           refineUsesAlignmentAndAA(GEP, GA, DL, AliasScope, NoAlias,
1521                                    MaxDepth - 1);
1522         }
1523         continue;
1524       }
1525       if (auto *I = dyn_cast<Instruction>(U)) {
1526         if (I->getOpcode() == Instruction::BitCast ||
1527             I->getOpcode() == Instruction::AddrSpaceCast)
1528           refineUsesAlignmentAndAA(I, A, DL, AliasScope, NoAlias, MaxDepth - 1);
1529       }
1530     }
1531   }
1532 };
1533 
1534 class AMDGPULowerModuleLDSLegacy : public ModulePass {
1535 public:
1536   const AMDGPUTargetMachine *TM;
1537   static char ID;
1538 
1539   AMDGPULowerModuleLDSLegacy(const AMDGPUTargetMachine *TM_ = nullptr)
1540       : ModulePass(ID), TM(TM_) {
1541     initializeAMDGPULowerModuleLDSLegacyPass(*PassRegistry::getPassRegistry());
1542   }
1543 
1544   void getAnalysisUsage(AnalysisUsage &AU) const override {
1545     if (!TM)
1546       AU.addRequired<TargetPassConfig>();
1547   }
1548 
1549   bool runOnModule(Module &M) override {
1550     if (!TM) {
1551       auto &TPC = getAnalysis<TargetPassConfig>();
1552       TM = &TPC.getTM<AMDGPUTargetMachine>();
1553     }
1554 
1555     return AMDGPULowerModuleLDS(*TM).runOnModule(M);
1556   }
1557 };
1558 
1559 } // namespace
1560 char AMDGPULowerModuleLDSLegacy::ID = 0;
1561 
1562 char &llvm::AMDGPULowerModuleLDSLegacyPassID = AMDGPULowerModuleLDSLegacy::ID;
1563 
1564 INITIALIZE_PASS_BEGIN(AMDGPULowerModuleLDSLegacy, DEBUG_TYPE,
1565                       "Lower uses of LDS variables from non-kernel functions",
1566                       false, false)
1567 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
1568 INITIALIZE_PASS_END(AMDGPULowerModuleLDSLegacy, DEBUG_TYPE,
1569                     "Lower uses of LDS variables from non-kernel functions",
1570                     false, false)
1571 
1572 ModulePass *
1573 llvm::createAMDGPULowerModuleLDSLegacyPass(const AMDGPUTargetMachine *TM) {
1574   return new AMDGPULowerModuleLDSLegacy(TM);
1575 }
1576 
1577 PreservedAnalyses AMDGPULowerModuleLDSPass::run(Module &M,
1578                                                 ModuleAnalysisManager &) {
1579   return AMDGPULowerModuleLDS(TM).runOnModule(M) ? PreservedAnalyses::none()
1580                                                  : PreservedAnalyses::all();
1581 }
1582