xref: /llvm-project/llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp (revision 125b90749a98d6dc6b492883c9617f9e91ab60e0)
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 // Kernels that do not use this struct are annoteated with the attribute
77 // amdgpu-elide-module-lds which allows the back end to elide the allocation.
78 //
79 // The "table" lowering implemented here has three components.
80 // First kernels are assigned a unique integer identifier which is available in
81 // functions it calls through the intrinsic amdgcn_lds_kernel_id. The integer
82 // is passed through a specific SGPR, thus works with indirect calls.
83 // Second, each kernel allocates LDS variables independent of other kernels and
84 // writes the addresses it chose for each variable into an array in consistent
85 // order. If the kernel does not allocate a given variable, it writes undef to
86 // the corresponding array location. These arrays are written to a constant
87 // table in the order matching the kernel unique integer identifier.
88 // Third, uses from non-kernel functions are replaced with a table lookup using
89 // the intrinsic function to find the address of the variable.
90 //
91 // "Kernel" lowering is only applicable for variables that are unambiguously
92 // reachable from exactly one kernel. For those cases, accesses to the variable
93 // can be lowered to ConstantExpr address of a struct instance specific to that
94 // one kernel. This is zero cost in space and in compute. It will raise a fatal
95 // error on any variable that might be reachable from multiple kernels and is
96 // thus most easily used as part of the hybrid lowering strategy.
97 //
98 // Hybrid lowering is a mixture of the above. It uses the zero cost kernel
99 // lowering where it can. It lowers the variable accessed by the greatest
100 // number of kernels using the module strategy as that is free for the first
101 // variable. Any futher variables that can be lowered with the module strategy
102 // without incurring LDS memory overhead are. The remaining ones are lowered
103 // via table.
104 //
105 // Consequences
106 // - No heuristics or user controlled magic numbers, hybrid is the right choice
107 // - Kernels that don't use functions (or have had them all inlined) are not
108 //   affected by any lowering for kernels that do.
109 // - Kernels that don't make indirect function calls are not affected by those
110 //   that do.
111 // - Variables which are used by lots of kernels, e.g. those injected by a
112 //   language runtime in most kernels, are expected to have no overhead
113 // - Implementations that instantiate templates per-kernel where those templates
114 //   use LDS are expected to hit the "Kernel" lowering strategy
115 // - The runtime properties impose a cost in compiler implementation complexity
116 //
117 // Dynamic LDS implementation
118 // Dynamic LDS is lowered similarly to the "table" strategy above and uses the
119 // same intrinsic to identify which kernel is at the root of the dynamic call
120 // graph. This relies on the specified behaviour that all dynamic LDS variables
121 // alias one another, i.e. are at the same address, with respect to a given
122 // kernel. Therefore this pass creates new dynamic LDS variables for each kernel
123 // that allocates any dynamic LDS and builds a table of addresses out of those.
124 // The AMDGPUPromoteAlloca pass skips kernels that use dynamic LDS.
125 // The corresponding optimisation for "kernel" lowering where the table lookup
126 // is elided is not implemented.
127 //
128 //
129 // Implementation notes / limitations
130 // A single LDS global variable represents an instance per kernel that can reach
131 // said variables. This pass essentially specialises said variables per kernel.
132 // Handling ConstantExpr during the pass complicated this significantly so now
133 // all ConstantExpr uses of LDS variables are expanded to instructions. This
134 // may need amending when implementing non-undef initialisers.
135 //
136 // Lowering is split between this IR pass and the back end. This pass chooses
137 // where given variables should be allocated and marks them with metadata,
138 // MD_absolute_symbol. The backend places the variables in coincidentally the
139 // same location and raises a fatal error if something has gone awry. This works
140 // in practice because the only pass between this one and the backend that
141 // changes LDS is PromoteAlloca and the changes it makes do not conflict.
142 //
143 // Addresses are written to constant global arrays based on the same metadata.
144 //
145 // The backend lowers LDS variables in the order of traversal of the function.
146 // This is at odds with the deterministic layout required. The workaround is to
147 // allocate the fixed-address variables immediately upon starting the function
148 // where they can be placed as intended. This requires a means of mapping from
149 // the function to the variables that it allocates. For the module scope lds,
150 // this is via metadata indicating whether the variable is not required. If a
151 // pass deletes that metadata, a fatal error on disagreement with the absolute
152 // symbol metadata will occur. For kernel scope and dynamic, this is by _name_
153 // correspondence between the function and the variable. It requires the
154 // kernel to have a name (which is only a limitation for tests in practice) and
155 // for nothing to rename the corresponding symbols. This is a hazard if the pass
156 // is run multiple times during debugging. Alternative schemes considered all
157 // involve bespoke metadata.
158 //
159 // If the name correspondence can be replaced, multiple distinct kernels that
160 // have the same memory layout can map to the same kernel id (as the address
161 // itself is handled by the absolute symbol metadata) and that will allow more
162 // uses of the "kernel" style faster lowering and reduce the size of the lookup
163 // tables.
164 //
165 // There is a test that checks this does not fire for a graphics shader. This
166 // lowering is expected to work for graphics if the isKernel test is changed.
167 //
168 // The current markUsedByKernel is sufficient for PromoteAlloca but is elided
169 // before codegen. Replacing this with an equivalent intrinsic which lasts until
170 // shortly after the machine function lowering of LDS would help break the name
171 // mapping. The other part needed is probably to amend PromoteAlloca to embed
172 // the LDS variables it creates in the same struct created here. That avoids the
173 // current hazard where a PromoteAlloca LDS variable might be allocated before
174 // the kernel scope (and thus error on the address check). Given a new invariant
175 // that no LDS variables exist outside of the structs managed here, and an
176 // intrinsic that lasts until after the LDS frame lowering, it should be
177 // possible to drop the name mapping and fold equivalent memory layouts.
178 //
179 //===----------------------------------------------------------------------===//
180 
181 #include "AMDGPU.h"
182 #include "Utils/AMDGPUBaseInfo.h"
183 #include "Utils/AMDGPUMemoryUtils.h"
184 #include "llvm/ADT/BitVector.h"
185 #include "llvm/ADT/DenseMap.h"
186 #include "llvm/ADT/DenseSet.h"
187 #include "llvm/ADT/STLExtras.h"
188 #include "llvm/ADT/SetOperations.h"
189 #include "llvm/ADT/SetVector.h"
190 #include "llvm/Analysis/CallGraph.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/OptimizedStructLayout.h"
204 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
205 #include "llvm/Transforms/Utils/ModuleUtils.h"
206 
207 #include <tuple>
208 #include <vector>
209 
210 #include <cstdio>
211 
212 #define DEBUG_TYPE "amdgpu-lower-module-lds"
213 
214 using namespace llvm;
215 
216 namespace {
217 
218 cl::opt<bool> SuperAlignLDSGlobals(
219     "amdgpu-super-align-lds-globals",
220     cl::desc("Increase alignment of LDS if it is not on align boundary"),
221     cl::init(true), cl::Hidden);
222 
223 enum class LoweringKind { module, table, kernel, hybrid };
224 cl::opt<LoweringKind> LoweringKindLoc(
225     "amdgpu-lower-module-lds-strategy",
226     cl::desc("Specify lowering strategy for function LDS access:"), cl::Hidden,
227     cl::init(LoweringKind::hybrid),
228     cl::values(
229         clEnumValN(LoweringKind::table, "table", "Lower via table lookup"),
230         clEnumValN(LoweringKind::module, "module", "Lower via module struct"),
231         clEnumValN(
232             LoweringKind::kernel, "kernel",
233             "Lower variables reachable from one kernel, otherwise abort"),
234         clEnumValN(LoweringKind::hybrid, "hybrid",
235                    "Lower via mixture of above strategies")));
236 
237 bool isKernelLDS(const Function *F) {
238   // Some weirdness here. AMDGPU::isKernelCC does not call into
239   // AMDGPU::isKernel with the calling conv, it instead calls into
240   // isModuleEntryFunction which returns true for more calling conventions
241   // than AMDGPU::isKernel does. There's a FIXME on AMDGPU::isKernel.
242   // There's also a test that checks that the LDS lowering does not hit on
243   // a graphics shader, denoted amdgpu_ps, so stay with the limited case.
244   // Putting LDS in the name of the function to draw attention to this.
245   return AMDGPU::isKernel(F->getCallingConv());
246 }
247 
248 template <typename T> std::vector<T> sortByName(std::vector<T> &&V) {
249   llvm::sort(V.begin(), V.end(), [](const auto *L, const auto *R) {
250     return L->getName() < R->getName();
251   });
252   return V;
253 }
254 
255 class AMDGPULowerModuleLDS : public ModulePass {
256 
257   static void
258   removeLocalVarsFromUsedLists(Module &M,
259                                const DenseSet<GlobalVariable *> &LocalVars) {
260     // The verifier rejects used lists containing an inttoptr of a constant
261     // so remove the variables from these lists before replaceAllUsesWith
262     SmallPtrSet<Constant *, 8> LocalVarsSet;
263     for (GlobalVariable *LocalVar : LocalVars)
264       LocalVarsSet.insert(cast<Constant>(LocalVar->stripPointerCasts()));
265 
266     removeFromUsedLists(
267         M, [&LocalVarsSet](Constant *C) { return LocalVarsSet.count(C); });
268 
269     for (GlobalVariable *LocalVar : LocalVars)
270       LocalVar->removeDeadConstantUsers();
271   }
272 
273   static void markUsedByKernel(IRBuilder<> &Builder, Function *Func,
274                                GlobalVariable *SGV) {
275     // The llvm.amdgcn.module.lds instance is implicitly used by all kernels
276     // that might call a function which accesses a field within it. This is
277     // presently approximated to 'all kernels' if there are any such functions
278     // in the module. This implicit use is redefined as an explicit use here so
279     // that later passes, specifically PromoteAlloca, account for the required
280     // memory without any knowledge of this transform.
281 
282     // An operand bundle on llvm.donothing works because the call instruction
283     // survives until after the last pass that needs to account for LDS. It is
284     // better than inline asm as the latter survives until the end of codegen. A
285     // totally robust solution would be a function with the same semantics as
286     // llvm.donothing that takes a pointer to the instance and is lowered to a
287     // no-op after LDS is allocated, but that is not presently necessary.
288 
289     // This intrinsic is eliminated shortly before instruction selection. It
290     // does not suffice to indicate to ISel that a given global which is not
291     // immediately used by the kernel must still be allocated by it. An
292     // equivalent target specific intrinsic which lasts until immediately after
293     // codegen would suffice for that, but one would still need to ensure that
294     // the variables are allocated in the anticpated order.
295 
296     LLVMContext &Ctx = Func->getContext();
297 
298     Builder.SetInsertPoint(Func->getEntryBlock().getFirstNonPHI());
299 
300     FunctionType *FTy = FunctionType::get(Type::getVoidTy(Ctx), {});
301 
302     Function *Decl =
303         Intrinsic::getDeclaration(Func->getParent(), Intrinsic::donothing, {});
304 
305     Value *UseInstance[1] = {Builder.CreateInBoundsGEP(
306         SGV->getValueType(), SGV, ConstantInt::get(Type::getInt32Ty(Ctx), 0))};
307 
308     Builder.CreateCall(FTy, Decl, {},
309                        {OperandBundleDefT<Value *>("ExplicitUse", UseInstance)},
310                        "");
311   }
312 
313   static bool eliminateConstantExprUsesOfLDSFromAllInstructions(Module &M) {
314     // Constants are uniqued within LLVM. A ConstantExpr referring to a LDS
315     // global may have uses from multiple different functions as a result.
316     // This pass specialises LDS variables with respect to the kernel that
317     // allocates them.
318 
319     // This is semantically equivalent to (the unimplemented as slow):
320     // for (auto &F : M.functions())
321     //   for (auto &BB : F)
322     //     for (auto &I : BB)
323     //       for (Use &Op : I.operands())
324     //         if (constantExprUsesLDS(Op))
325     //           replaceConstantExprInFunction(I, Op);
326 
327     SmallVector<Constant *> LDSGlobals;
328     for (auto &GV : M.globals())
329       if (AMDGPU::isLDSVariableToLower(GV))
330         LDSGlobals.push_back(&GV);
331 
332     return convertUsersOfConstantsToInstructions(LDSGlobals);
333   }
334 
335 public:
336   static char ID;
337 
338   AMDGPULowerModuleLDS() : ModulePass(ID) {
339     initializeAMDGPULowerModuleLDSPass(*PassRegistry::getPassRegistry());
340   }
341 
342   using FunctionVariableMap = DenseMap<Function *, DenseSet<GlobalVariable *>>;
343 
344   using VariableFunctionMap = DenseMap<GlobalVariable *, DenseSet<Function *>>;
345 
346   static void getUsesOfLDSByFunction(CallGraph const &CG, Module &M,
347                                      FunctionVariableMap &kernels,
348                                      FunctionVariableMap &functions) {
349 
350     // Get uses from the current function, excluding uses by called functions
351     // Two output variables to avoid walking the globals list twice
352     for (auto &GV : M.globals()) {
353       if (!AMDGPU::isLDSVariableToLower(GV)) {
354         continue;
355       }
356 
357       for (User *V : GV.users()) {
358         if (auto *I = dyn_cast<Instruction>(V)) {
359           Function *F = I->getFunction();
360           if (isKernelLDS(F)) {
361             kernels[F].insert(&GV);
362           } else {
363             functions[F].insert(&GV);
364           }
365         }
366       }
367     }
368   }
369 
370   struct LDSUsesInfoTy {
371     FunctionVariableMap direct_access;
372     FunctionVariableMap indirect_access;
373   };
374 
375   static LDSUsesInfoTy getTransitiveUsesOfLDS(CallGraph const &CG, Module &M) {
376 
377     FunctionVariableMap direct_map_kernel;
378     FunctionVariableMap direct_map_function;
379     getUsesOfLDSByFunction(CG, M, direct_map_kernel, direct_map_function);
380 
381     // Collect variables that are used by functions whose address has escaped
382     DenseSet<GlobalVariable *> VariablesReachableThroughFunctionPointer;
383     for (Function &F : M.functions()) {
384       if (!isKernelLDS(&F))
385         if (F.hasAddressTaken(nullptr,
386                               /* IgnoreCallbackUses */ false,
387                               /* IgnoreAssumeLikeCalls */ false,
388                               /* IgnoreLLVMUsed */ true,
389                               /* IgnoreArcAttachedCall */ false)) {
390           set_union(VariablesReachableThroughFunctionPointer,
391                     direct_map_function[&F]);
392         }
393     }
394 
395     auto functionMakesUnknownCall = [&](const Function *F) -> bool {
396       assert(!F->isDeclaration());
397       for (const CallGraphNode::CallRecord &R : *CG[F]) {
398         if (!R.second->getFunction()) {
399           return true;
400         }
401       }
402       return false;
403     };
404 
405     // Work out which variables are reachable through function calls
406     FunctionVariableMap transitive_map_function = direct_map_function;
407 
408     // If the function makes any unknown call, assume the worst case that it can
409     // access all variables accessed by functions whose address escaped
410     for (Function &F : M.functions()) {
411       if (!F.isDeclaration() && functionMakesUnknownCall(&F)) {
412         if (!isKernelLDS(&F)) {
413           set_union(transitive_map_function[&F],
414                     VariablesReachableThroughFunctionPointer);
415         }
416       }
417     }
418 
419     // Direct implementation of collecting all variables reachable from each
420     // function
421     for (Function &Func : M.functions()) {
422       if (Func.isDeclaration() || isKernelLDS(&Func))
423         continue;
424 
425       DenseSet<Function *> seen; // catches cycles
426       SmallVector<Function *, 4> wip{&Func};
427 
428       while (!wip.empty()) {
429         Function *F = wip.pop_back_val();
430 
431         // Can accelerate this by referring to transitive map for functions that
432         // have already been computed, with more care than this
433         set_union(transitive_map_function[&Func], direct_map_function[F]);
434 
435         for (const CallGraphNode::CallRecord &R : *CG[F]) {
436           Function *ith = R.second->getFunction();
437           if (ith) {
438             if (!seen.contains(ith)) {
439               seen.insert(ith);
440               wip.push_back(ith);
441             }
442           }
443         }
444       }
445     }
446 
447     // direct_map_kernel lists which variables are used by the kernel
448     // find the variables which are used through a function call
449     FunctionVariableMap indirect_map_kernel;
450 
451     for (Function &Func : M.functions()) {
452       if (Func.isDeclaration() || !isKernelLDS(&Func))
453         continue;
454 
455       for (const CallGraphNode::CallRecord &R : *CG[&Func]) {
456         Function *ith = R.second->getFunction();
457         if (ith) {
458           set_union(indirect_map_kernel[&Func], transitive_map_function[ith]);
459         } else {
460           set_union(indirect_map_kernel[&Func],
461                     VariablesReachableThroughFunctionPointer);
462         }
463       }
464     }
465 
466     return {std::move(direct_map_kernel), std::move(indirect_map_kernel)};
467   }
468 
469   struct LDSVariableReplacement {
470     GlobalVariable *SGV = nullptr;
471     DenseMap<GlobalVariable *, Constant *> LDSVarsToConstantGEP;
472   };
473 
474   // remap from lds global to a constantexpr gep to where it has been moved to
475   // for each kernel
476   // an array with an element for each kernel containing where the corresponding
477   // variable was remapped to
478 
479   static Constant *getAddressesOfVariablesInKernel(
480       LLVMContext &Ctx, ArrayRef<GlobalVariable *> Variables,
481       const DenseMap<GlobalVariable *, Constant *> &LDSVarsToConstantGEP) {
482     // Create a ConstantArray containing the address of each Variable within the
483     // kernel corresponding to LDSVarsToConstantGEP, or poison if that kernel
484     // does not allocate it
485     // TODO: Drop the ptrtoint conversion
486 
487     Type *I32 = Type::getInt32Ty(Ctx);
488 
489     ArrayType *KernelOffsetsType = ArrayType::get(I32, Variables.size());
490 
491     SmallVector<Constant *> Elements;
492     for (size_t i = 0; i < Variables.size(); i++) {
493       GlobalVariable *GV = Variables[i];
494       auto ConstantGepIt = LDSVarsToConstantGEP.find(GV);
495       if (ConstantGepIt != LDSVarsToConstantGEP.end()) {
496         auto elt = ConstantExpr::getPtrToInt(ConstantGepIt->second, I32);
497         Elements.push_back(elt);
498       } else {
499         Elements.push_back(PoisonValue::get(I32));
500       }
501     }
502     return ConstantArray::get(KernelOffsetsType, Elements);
503   }
504 
505   static GlobalVariable *buildLookupTable(
506       Module &M, ArrayRef<GlobalVariable *> Variables,
507       ArrayRef<Function *> kernels,
508       DenseMap<Function *, LDSVariableReplacement> &KernelToReplacement) {
509     if (Variables.empty()) {
510       return nullptr;
511     }
512     LLVMContext &Ctx = M.getContext();
513 
514     const size_t NumberVariables = Variables.size();
515     const size_t NumberKernels = kernels.size();
516 
517     ArrayType *KernelOffsetsType =
518         ArrayType::get(Type::getInt32Ty(Ctx), NumberVariables);
519 
520     ArrayType *AllKernelsOffsetsType =
521         ArrayType::get(KernelOffsetsType, NumberKernels);
522 
523     std::vector<Constant *> overallConstantExprElts(NumberKernels);
524     for (size_t i = 0; i < NumberKernels; i++) {
525       LDSVariableReplacement Replacement = KernelToReplacement[kernels[i]];
526       overallConstantExprElts[i] = getAddressesOfVariablesInKernel(
527           Ctx, Variables, Replacement.LDSVarsToConstantGEP);
528     }
529 
530     Constant *init =
531         ConstantArray::get(AllKernelsOffsetsType, overallConstantExprElts);
532 
533     return new GlobalVariable(
534         M, AllKernelsOffsetsType, true, GlobalValue::InternalLinkage, init,
535         "llvm.amdgcn.lds.offset.table", nullptr, GlobalValue::NotThreadLocal,
536         AMDGPUAS::CONSTANT_ADDRESS);
537   }
538 
539   void replaceUseWithTableLookup(Module &M, IRBuilder<> &Builder,
540                                  GlobalVariable *LookupTable,
541                                  GlobalVariable *GV, Use &U,
542                                  Value *OptionalIndex) {
543     // Table is a constant array of the same length as OrderedKernels
544     LLVMContext &Ctx = M.getContext();
545     Type *I32 = Type::getInt32Ty(Ctx);
546     auto *I = cast<Instruction>(U.getUser());
547 
548     Value *tableKernelIndex = getTableLookupKernelIndex(M, I->getFunction());
549 
550     if (auto *Phi = dyn_cast<PHINode>(I)) {
551       BasicBlock *BB = Phi->getIncomingBlock(U);
552       Builder.SetInsertPoint(&(*(BB->getFirstInsertionPt())));
553     } else {
554       Builder.SetInsertPoint(I);
555     }
556 
557     SmallVector<Value *, 3> GEPIdx = {
558         ConstantInt::get(I32, 0),
559         tableKernelIndex,
560     };
561     if (OptionalIndex)
562       GEPIdx.push_back(OptionalIndex);
563 
564     Value *Address = Builder.CreateInBoundsGEP(
565         LookupTable->getValueType(), LookupTable, GEPIdx, GV->getName());
566 
567     Value *loaded = Builder.CreateLoad(I32, Address);
568 
569     Value *replacement =
570         Builder.CreateIntToPtr(loaded, GV->getType(), GV->getName());
571 
572     U.set(replacement);
573   }
574 
575   void replaceUsesInInstructionsWithTableLookup(
576       Module &M, ArrayRef<GlobalVariable *> ModuleScopeVariables,
577       GlobalVariable *LookupTable) {
578 
579     LLVMContext &Ctx = M.getContext();
580     IRBuilder<> Builder(Ctx);
581     Type *I32 = Type::getInt32Ty(Ctx);
582 
583     for (size_t Index = 0; Index < ModuleScopeVariables.size(); Index++) {
584       auto *GV = ModuleScopeVariables[Index];
585 
586       for (Use &U : make_early_inc_range(GV->uses())) {
587         auto *I = dyn_cast<Instruction>(U.getUser());
588         if (!I)
589           continue;
590 
591         replaceUseWithTableLookup(M, Builder, LookupTable, GV, U,
592                                   ConstantInt::get(I32, Index));
593       }
594     }
595   }
596 
597   static DenseSet<Function *> kernelsThatIndirectlyAccessAnyOfPassedVariables(
598       Module &M, LDSUsesInfoTy &LDSUsesInfo,
599       DenseSet<GlobalVariable *> const &VariableSet) {
600 
601     DenseSet<Function *> KernelSet;
602 
603     if (VariableSet.empty())
604       return KernelSet;
605 
606     for (Function &Func : M.functions()) {
607       if (Func.isDeclaration() || !isKernelLDS(&Func))
608         continue;
609       for (GlobalVariable *GV : LDSUsesInfo.indirect_access[&Func]) {
610         if (VariableSet.contains(GV)) {
611           KernelSet.insert(&Func);
612           break;
613         }
614       }
615     }
616 
617     return KernelSet;
618   }
619 
620   static GlobalVariable *
621   chooseBestVariableForModuleStrategy(const DataLayout &DL,
622                                       VariableFunctionMap &LDSVars) {
623     // Find the global variable with the most indirect uses from kernels
624 
625     struct CandidateTy {
626       GlobalVariable *GV = nullptr;
627       size_t UserCount = 0;
628       size_t Size = 0;
629 
630       CandidateTy() = default;
631 
632       CandidateTy(GlobalVariable *GV, uint64_t UserCount, uint64_t AllocSize)
633           : GV(GV), UserCount(UserCount), Size(AllocSize) {}
634 
635       bool operator<(const CandidateTy &Other) const {
636         // Fewer users makes module scope variable less attractive
637         if (UserCount < Other.UserCount) {
638           return true;
639         }
640         if (UserCount > Other.UserCount) {
641           return false;
642         }
643 
644         // Bigger makes module scope variable less attractive
645         if (Size < Other.Size) {
646           return false;
647         }
648 
649         if (Size > Other.Size) {
650           return true;
651         }
652 
653         // Arbitrary but consistent
654         return GV->getName() < Other.GV->getName();
655       }
656     };
657 
658     CandidateTy MostUsed;
659 
660     for (auto &K : LDSVars) {
661       GlobalVariable *GV = K.first;
662       if (K.second.size() <= 1) {
663         // A variable reachable by only one kernel is best lowered with kernel
664         // strategy
665         continue;
666       }
667       CandidateTy Candidate(
668           GV, K.second.size(),
669           DL.getTypeAllocSize(GV->getValueType()).getFixedValue());
670       if (MostUsed < Candidate)
671         MostUsed = Candidate;
672     }
673 
674     return MostUsed.GV;
675   }
676 
677   static void recordLDSAbsoluteAddress(Module *M, GlobalVariable *GV,
678                                        uint32_t Address) {
679     // Write the specified address into metadata where it can be retrieved by
680     // the assembler. Format is a half open range, [Address Address+1)
681     LLVMContext &Ctx = M->getContext();
682     auto *IntTy =
683         M->getDataLayout().getIntPtrType(Ctx, AMDGPUAS::LOCAL_ADDRESS);
684     auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntTy, Address));
685     auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntTy, Address + 1));
686     GV->setMetadata(LLVMContext::MD_absolute_symbol,
687                     MDNode::get(Ctx, {MinC, MaxC}));
688   }
689 
690   DenseMap<Function *, Value *> tableKernelIndexCache;
691   Value *getTableLookupKernelIndex(Module &M, Function *F) {
692     // Accesses from a function use the amdgcn_lds_kernel_id intrinsic which
693     // lowers to a read from a live in register. Emit it once in the entry
694     // block to spare deduplicating it later.
695     if (tableKernelIndexCache.count(F) == 0) {
696       LLVMContext &Ctx = M.getContext();
697       IRBuilder<> Builder(Ctx);
698       FunctionType *FTy = FunctionType::get(Type::getInt32Ty(Ctx), {});
699       Function *Decl =
700           Intrinsic::getDeclaration(&M, Intrinsic::amdgcn_lds_kernel_id, {});
701 
702       BasicBlock::iterator it =
703           F->getEntryBlock().getFirstNonPHIOrDbgOrAlloca();
704       Instruction &i = *it;
705       Builder.SetInsertPoint(&i);
706 
707       tableKernelIndexCache[F] = Builder.CreateCall(FTy, Decl, {});
708     }
709 
710     return tableKernelIndexCache[F];
711   }
712 
713   static std::vector<Function *> assignLDSKernelIDToEachKernel(
714       Module *M, DenseSet<Function *> const &KernelsThatAllocateTableLDS,
715       DenseSet<Function *> const &KernelsThatIndirectlyAllocateDynamicLDS) {
716     // Associate kernels in the set with an arbirary but reproducible order and
717     // annotate them with that order in metadata. This metadata is recognised by
718     // the backend and lowered to a SGPR which can be read from using
719     // amdgcn_lds_kernel_id.
720 
721     std::vector<Function *> OrderedKernels;
722     if (!KernelsThatAllocateTableLDS.empty() ||
723         !KernelsThatIndirectlyAllocateDynamicLDS.empty()) {
724 
725       for (Function &Func : M->functions()) {
726         if (Func.isDeclaration())
727           continue;
728         if (!isKernelLDS(&Func))
729           continue;
730 
731         if (KernelsThatAllocateTableLDS.contains(&Func) ||
732             KernelsThatIndirectlyAllocateDynamicLDS.contains(&Func)) {
733           assert(Func.hasName()); // else fatal error earlier
734           OrderedKernels.push_back(&Func);
735         }
736       }
737 
738       // Put them in an arbitrary but reproducible order
739       OrderedKernels = sortByName(std::move(OrderedKernels));
740 
741       // Annotate the kernels with their order in this vector
742       LLVMContext &Ctx = M->getContext();
743       IRBuilder<> Builder(Ctx);
744 
745       if (OrderedKernels.size() > UINT32_MAX) {
746         // 32 bit keeps it in one SGPR. > 2**32 kernels won't fit on the GPU
747         report_fatal_error("Unimplemented LDS lowering for > 2**32 kernels");
748       }
749 
750       for (size_t i = 0; i < OrderedKernels.size(); i++) {
751         Metadata *AttrMDArgs[1] = {
752             ConstantAsMetadata::get(Builder.getInt32(i)),
753         };
754         OrderedKernels[i]->setMetadata("llvm.amdgcn.lds.kernel.id",
755                                        MDNode::get(Ctx, AttrMDArgs));
756       }
757     }
758     return OrderedKernels;
759   }
760 
761   static void partitionVariablesIntoIndirectStrategies(
762       Module &M, LDSUsesInfoTy const &LDSUsesInfo,
763       VariableFunctionMap &LDSToKernelsThatNeedToAccessItIndirectly,
764       DenseSet<GlobalVariable *> &ModuleScopeVariables,
765       DenseSet<GlobalVariable *> &TableLookupVariables,
766       DenseSet<GlobalVariable *> &KernelAccessVariables,
767       DenseSet<GlobalVariable *> &DynamicVariables) {
768 
769     GlobalVariable *HybridModuleRoot =
770         LoweringKindLoc != LoweringKind::hybrid
771             ? nullptr
772             : chooseBestVariableForModuleStrategy(
773                   M.getDataLayout(), LDSToKernelsThatNeedToAccessItIndirectly);
774 
775     DenseSet<Function *> const EmptySet;
776     DenseSet<Function *> const &HybridModuleRootKernels =
777         HybridModuleRoot
778             ? LDSToKernelsThatNeedToAccessItIndirectly[HybridModuleRoot]
779             : EmptySet;
780 
781     for (auto &K : LDSToKernelsThatNeedToAccessItIndirectly) {
782       // Each iteration of this loop assigns exactly one global variable to
783       // exactly one of the implementation strategies.
784 
785       GlobalVariable *GV = K.first;
786       assert(AMDGPU::isLDSVariableToLower(*GV));
787       assert(K.second.size() != 0);
788 
789       if (AMDGPU::isDynamicLDS(*GV)) {
790         DynamicVariables.insert(GV);
791         continue;
792       }
793 
794       switch (LoweringKindLoc) {
795       case LoweringKind::module:
796         ModuleScopeVariables.insert(GV);
797         break;
798 
799       case LoweringKind::table:
800         TableLookupVariables.insert(GV);
801         break;
802 
803       case LoweringKind::kernel:
804         if (K.second.size() == 1) {
805           KernelAccessVariables.insert(GV);
806         } else {
807           report_fatal_error(
808               "cannot lower LDS '" + GV->getName() +
809               "' to kernel access as it is reachable from multiple kernels");
810         }
811         break;
812 
813       case LoweringKind::hybrid: {
814         if (GV == HybridModuleRoot) {
815           assert(K.second.size() != 1);
816           ModuleScopeVariables.insert(GV);
817         } else if (K.second.size() == 1) {
818           KernelAccessVariables.insert(GV);
819         } else if (set_is_subset(K.second, HybridModuleRootKernels)) {
820           ModuleScopeVariables.insert(GV);
821         } else {
822           TableLookupVariables.insert(GV);
823         }
824         break;
825       }
826       }
827     }
828 
829     // All LDS variables accessed indirectly have now been partitioned into
830     // the distinct lowering strategies.
831     assert(ModuleScopeVariables.size() + TableLookupVariables.size() +
832                KernelAccessVariables.size() + DynamicVariables.size() ==
833            LDSToKernelsThatNeedToAccessItIndirectly.size());
834   }
835 
836   static GlobalVariable *lowerModuleScopeStructVariables(
837       Module &M, DenseSet<GlobalVariable *> const &ModuleScopeVariables,
838       DenseSet<Function *> const &KernelsThatAllocateModuleLDS) {
839     // Create a struct to hold the ModuleScopeVariables
840     // Replace all uses of those variables from non-kernel functions with the
841     // new struct instance Replace only the uses from kernel functions that will
842     // allocate this instance. That is a space optimisation - kernels that use a
843     // subset of the module scope struct and do not need to allocate it for
844     // indirect calls will only allocate the subset they use (they do so as part
845     // of the per-kernel lowering).
846     if (ModuleScopeVariables.empty()) {
847       return nullptr;
848     }
849 
850     LLVMContext &Ctx = M.getContext();
851 
852     LDSVariableReplacement ModuleScopeReplacement =
853         createLDSVariableReplacement(M, "llvm.amdgcn.module.lds",
854                                      ModuleScopeVariables);
855 
856     appendToCompilerUsed(M, {static_cast<GlobalValue *>(
857                                 ConstantExpr::getPointerBitCastOrAddrSpaceCast(
858                                     cast<Constant>(ModuleScopeReplacement.SGV),
859                                     Type::getInt8PtrTy(Ctx)))});
860 
861     // module.lds will be allocated at zero in any kernel that allocates it
862     recordLDSAbsoluteAddress(&M, ModuleScopeReplacement.SGV, 0);
863 
864     // historic
865     removeLocalVarsFromUsedLists(M, ModuleScopeVariables);
866 
867     // Replace all uses of module scope variable from non-kernel functions
868     replaceLDSVariablesWithStruct(
869         M, ModuleScopeVariables, ModuleScopeReplacement, [&](Use &U) {
870           Instruction *I = dyn_cast<Instruction>(U.getUser());
871           if (!I) {
872             return false;
873           }
874           Function *F = I->getFunction();
875           return !isKernelLDS(F);
876         });
877 
878     // Replace uses of module scope variable from kernel functions that
879     // allocate the module scope variable, otherwise leave them unchanged
880     // Record on each kernel whether the module scope global is used by it
881 
882     IRBuilder<> Builder(Ctx);
883 
884     for (Function &Func : M.functions()) {
885       if (Func.isDeclaration() || !isKernelLDS(&Func))
886         continue;
887 
888       if (KernelsThatAllocateModuleLDS.contains(&Func)) {
889         replaceLDSVariablesWithStruct(
890             M, ModuleScopeVariables, ModuleScopeReplacement, [&](Use &U) {
891               Instruction *I = dyn_cast<Instruction>(U.getUser());
892               if (!I) {
893                 return false;
894               }
895               Function *F = I->getFunction();
896               return F == &Func;
897             });
898 
899         markUsedByKernel(Builder, &Func, ModuleScopeReplacement.SGV);
900 
901       } else {
902         markElideModuleLDS(Func);
903       }
904     }
905 
906     return ModuleScopeReplacement.SGV;
907   }
908 
909   static DenseMap<Function *, LDSVariableReplacement>
910   lowerKernelScopeStructVariables(
911       Module &M, LDSUsesInfoTy &LDSUsesInfo,
912       DenseSet<GlobalVariable *> const &ModuleScopeVariables,
913       DenseSet<Function *> const &KernelsThatAllocateModuleLDS,
914       GlobalVariable *MaybeModuleScopeStruct) {
915 
916     // Create a struct for each kernel for the non-module-scope variables.
917 
918     DenseMap<Function *, LDSVariableReplacement> KernelToReplacement;
919     for (Function &Func : M.functions()) {
920       if (Func.isDeclaration() || !isKernelLDS(&Func))
921         continue;
922 
923       DenseSet<GlobalVariable *> KernelUsedVariables;
924       // Allocating variables that are used directly in this struct to get
925       // alignment aware allocation and predictable frame size.
926       for (auto &v : LDSUsesInfo.direct_access[&Func]) {
927         if (!AMDGPU::isDynamicLDS(*v)) {
928           KernelUsedVariables.insert(v);
929         }
930       }
931 
932       // Allocating variables that are accessed indirectly so that a lookup of
933       // this struct instance can find them from nested functions.
934       for (auto &v : LDSUsesInfo.indirect_access[&Func]) {
935         if (!AMDGPU::isDynamicLDS(*v)) {
936           KernelUsedVariables.insert(v);
937         }
938       }
939 
940       // Variables allocated in module lds must all resolve to that struct,
941       // not to the per-kernel instance.
942       if (KernelsThatAllocateModuleLDS.contains(&Func)) {
943         for (GlobalVariable *v : ModuleScopeVariables) {
944           KernelUsedVariables.erase(v);
945         }
946       }
947 
948       if (KernelUsedVariables.empty()) {
949         // Either used no LDS, or the LDS it used was all in the module struct
950         // or dynamically sized
951         continue;
952       }
953 
954       // The association between kernel function and LDS struct is done by
955       // symbol name, which only works if the function in question has a
956       // name This is not expected to be a problem in practice as kernels
957       // are called by name making anonymous ones (which are named by the
958       // backend) difficult to use. This does mean that llvm test cases need
959       // to name the kernels.
960       if (!Func.hasName()) {
961         report_fatal_error("Anonymous kernels cannot use LDS variables");
962       }
963 
964       std::string VarName =
965           (Twine("llvm.amdgcn.kernel.") + Func.getName() + ".lds").str();
966 
967       auto Replacement =
968           createLDSVariableReplacement(M, VarName, KernelUsedVariables);
969 
970       // remove preserves existing codegen
971       removeLocalVarsFromUsedLists(M, KernelUsedVariables);
972       KernelToReplacement[&Func] = Replacement;
973 
974       // Rewrite uses within kernel to the new struct
975       replaceLDSVariablesWithStruct(
976           M, KernelUsedVariables, Replacement, [&Func](Use &U) {
977             Instruction *I = dyn_cast<Instruction>(U.getUser());
978             return I && I->getFunction() == &Func;
979           });
980     }
981     return KernelToReplacement;
982   }
983 
984   static GlobalVariable *
985   buildRepresentativeDynamicLDSInstance(Module &M, LDSUsesInfoTy &LDSUsesInfo,
986                                         Function *func) {
987     // Create a dynamic lds variable with a name associated with the passed
988     // function that has the maximum alignment of any dynamic lds variable
989     // reachable from this kernel. Dynamic LDS is allocated after the static LDS
990     // allocation, possibly after alignment padding. The representative variable
991     // created here has the maximum alignment of any other dynamic variable
992     // reachable by that kernel. All dynamic LDS variables are allocated at the
993     // same address in each kernel in order to provide the documented aliasing
994     // semantics. Setting the alignment here allows this IR pass to accurately
995     // predict the exact constant at which it will be allocated.
996 
997     assert(isKernelLDS(func));
998 
999     LLVMContext &Ctx = M.getContext();
1000     const DataLayout &DL = M.getDataLayout();
1001     Align MaxDynamicAlignment(1);
1002 
1003     auto UpdateMaxAlignment = [&MaxDynamicAlignment, &DL](GlobalVariable *GV) {
1004       if (AMDGPU::isDynamicLDS(*GV)) {
1005         MaxDynamicAlignment =
1006             std::max(MaxDynamicAlignment, AMDGPU::getAlign(DL, GV));
1007       }
1008     };
1009 
1010     for (GlobalVariable *GV : LDSUsesInfo.indirect_access[func]) {
1011       UpdateMaxAlignment(GV);
1012     }
1013 
1014     for (GlobalVariable *GV : LDSUsesInfo.direct_access[func]) {
1015       UpdateMaxAlignment(GV);
1016     }
1017 
1018     assert(func->hasName()); // Checked by caller
1019     auto emptyCharArray = ArrayType::get(Type::getInt8Ty(Ctx), 0);
1020     GlobalVariable *N = new GlobalVariable(
1021         M, emptyCharArray, false, GlobalValue::ExternalLinkage, nullptr,
1022         Twine("llvm.amdgcn." + func->getName() + ".dynlds"), nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS,
1023         false);
1024     N->setAlignment(MaxDynamicAlignment);
1025 
1026     assert(AMDGPU::isDynamicLDS(*N));
1027     return N;
1028   }
1029 
1030   DenseMap<Function *, GlobalVariable *> lowerDynamicLDSVariables(
1031       Module &M, LDSUsesInfoTy &LDSUsesInfo,
1032       DenseSet<Function *> const &KernelsThatIndirectlyAllocateDynamicLDS,
1033       DenseSet<GlobalVariable *> const &DynamicVariables,
1034       std::vector<Function *> const &OrderedKernels) {
1035     DenseMap<Function *, GlobalVariable *> KernelToCreatedDynamicLDS;
1036     if (!KernelsThatIndirectlyAllocateDynamicLDS.empty()) {
1037       LLVMContext &Ctx = M.getContext();
1038       IRBuilder<> Builder(Ctx);
1039       Type *I32 = Type::getInt32Ty(Ctx);
1040 
1041       std::vector<Constant *> newDynamicLDS;
1042 
1043       // Table is built in the same order as OrderedKernels
1044       for (auto &func : OrderedKernels) {
1045 
1046         if (KernelsThatIndirectlyAllocateDynamicLDS.contains(func)) {
1047           assert(isKernelLDS(func));
1048           if (!func->hasName()) {
1049             report_fatal_error("Anonymous kernels cannot use LDS variables");
1050           }
1051 
1052           GlobalVariable *N =
1053               buildRepresentativeDynamicLDSInstance(M, LDSUsesInfo, func);
1054 
1055           KernelToCreatedDynamicLDS[func] = N;
1056 
1057           markUsedByKernel(Builder, func, N);
1058 
1059           auto emptyCharArray = ArrayType::get(Type::getInt8Ty(Ctx), 0);
1060           auto GEP = ConstantExpr::getGetElementPtr(
1061               emptyCharArray, N, ConstantInt::get(I32, 0), true);
1062           newDynamicLDS.push_back(ConstantExpr::getPtrToInt(GEP, I32));
1063         } else {
1064           newDynamicLDS.push_back(PoisonValue::get(I32));
1065         }
1066       }
1067       assert(OrderedKernels.size() == newDynamicLDS.size());
1068 
1069       ArrayType *t = ArrayType::get(I32, newDynamicLDS.size());
1070       Constant *init = ConstantArray::get(t, newDynamicLDS);
1071       GlobalVariable *table = new GlobalVariable(
1072           M, t, true, GlobalValue::InternalLinkage, init,
1073           "llvm.amdgcn.dynlds.offset.table", nullptr,
1074           GlobalValue::NotThreadLocal, AMDGPUAS::CONSTANT_ADDRESS);
1075 
1076       for (GlobalVariable *GV : DynamicVariables) {
1077         for (Use &U : make_early_inc_range(GV->uses())) {
1078           auto *I = dyn_cast<Instruction>(U.getUser());
1079           if (!I)
1080             continue;
1081           if (isKernelLDS(I->getFunction()))
1082             continue;
1083 
1084           replaceUseWithTableLookup(M, Builder, table, GV, U, nullptr);
1085         }
1086       }
1087     }
1088     return KernelToCreatedDynamicLDS;
1089   }
1090 
1091   static bool canElideModuleLDS(const Function &F) {
1092     return F.hasFnAttribute("amdgpu-elide-module-lds");
1093   }
1094 
1095   static void markElideModuleLDS(Function &F) {
1096     F.addFnAttr("amdgpu-elide-module-lds");
1097   }
1098 
1099   bool runOnModule(Module &M) override {
1100     CallGraph CG = CallGraph(M);
1101     bool Changed = superAlignLDSGlobals(M);
1102 
1103     Changed |= eliminateConstantExprUsesOfLDSFromAllInstructions(M);
1104 
1105     Changed = true; // todo: narrow this down
1106 
1107     // For each kernel, what variables does it access directly or through
1108     // callees
1109     LDSUsesInfoTy LDSUsesInfo = getTransitiveUsesOfLDS(CG, M);
1110 
1111     // For each variable accessed through callees, which kernels access it
1112     VariableFunctionMap LDSToKernelsThatNeedToAccessItIndirectly;
1113     for (auto &K : LDSUsesInfo.indirect_access) {
1114       Function *F = K.first;
1115       assert(isKernelLDS(F));
1116       for (GlobalVariable *GV : K.second) {
1117         LDSToKernelsThatNeedToAccessItIndirectly[GV].insert(F);
1118       }
1119     }
1120 
1121     // Partition variables accessed indirectly into the different strategies
1122     DenseSet<GlobalVariable *> ModuleScopeVariables;
1123     DenseSet<GlobalVariable *> TableLookupVariables;
1124     DenseSet<GlobalVariable *> KernelAccessVariables;
1125     DenseSet<GlobalVariable *> DynamicVariables;
1126     partitionVariablesIntoIndirectStrategies(
1127         M, LDSUsesInfo, LDSToKernelsThatNeedToAccessItIndirectly,
1128         ModuleScopeVariables, TableLookupVariables, KernelAccessVariables,
1129         DynamicVariables);
1130 
1131     // If the kernel accesses a variable that is going to be stored in the
1132     // module instance through a call then that kernel needs to allocate the
1133     // module instance
1134     const DenseSet<Function *> KernelsThatAllocateModuleLDS =
1135         kernelsThatIndirectlyAccessAnyOfPassedVariables(M, LDSUsesInfo,
1136                                                         ModuleScopeVariables);
1137     const DenseSet<Function *> KernelsThatAllocateTableLDS =
1138         kernelsThatIndirectlyAccessAnyOfPassedVariables(M, LDSUsesInfo,
1139                                                         TableLookupVariables);
1140 
1141     const DenseSet<Function *> KernelsThatIndirectlyAllocateDynamicLDS =
1142         kernelsThatIndirectlyAccessAnyOfPassedVariables(M, LDSUsesInfo,
1143                                                         DynamicVariables);
1144 
1145     GlobalVariable *MaybeModuleScopeStruct = lowerModuleScopeStructVariables(
1146         M, ModuleScopeVariables, KernelsThatAllocateModuleLDS);
1147 
1148     DenseMap<Function *, LDSVariableReplacement> KernelToReplacement =
1149         lowerKernelScopeStructVariables(M, LDSUsesInfo, ModuleScopeVariables,
1150                                         KernelsThatAllocateModuleLDS,
1151                                         MaybeModuleScopeStruct);
1152 
1153     // Lower zero cost accesses to the kernel instances just created
1154     for (auto &GV : KernelAccessVariables) {
1155       auto &funcs = LDSToKernelsThatNeedToAccessItIndirectly[GV];
1156       assert(funcs.size() == 1); // Only one kernel can access it
1157       LDSVariableReplacement Replacement =
1158           KernelToReplacement[*(funcs.begin())];
1159 
1160       DenseSet<GlobalVariable *> Vec;
1161       Vec.insert(GV);
1162 
1163       // TODO: Looks like a latent bug, Replacement may not be marked
1164       // UsedByKernel here
1165       replaceLDSVariablesWithStruct(M, Vec, Replacement, [](Use &U) {
1166         return isa<Instruction>(U.getUser());
1167       });
1168     }
1169 
1170     // The ith element of this vector is kernel id i
1171     std::vector<Function *> OrderedKernels =
1172         assignLDSKernelIDToEachKernel(&M, KernelsThatAllocateTableLDS,
1173                                       KernelsThatIndirectlyAllocateDynamicLDS);
1174 
1175     if (!KernelsThatAllocateTableLDS.empty()) {
1176       LLVMContext &Ctx = M.getContext();
1177       IRBuilder<> Builder(Ctx);
1178 
1179       for (size_t i = 0; i < OrderedKernels.size(); i++) {
1180         markUsedByKernel(Builder, OrderedKernels[i],
1181                          KernelToReplacement[OrderedKernels[i]].SGV);
1182       }
1183 
1184       // The order must be consistent between lookup table and accesses to
1185       // lookup table
1186       auto TableLookupVariablesOrdered = sortByName(std::vector(
1187           TableLookupVariables.begin(), TableLookupVariables.end()));
1188 
1189       GlobalVariable *LookupTable = buildLookupTable(
1190           M, TableLookupVariablesOrdered, OrderedKernels, KernelToReplacement);
1191       replaceUsesInInstructionsWithTableLookup(M, TableLookupVariablesOrdered,
1192                                                LookupTable);
1193     }
1194 
1195     DenseMap<Function *, GlobalVariable *> KernelToCreatedDynamicLDS =
1196         lowerDynamicLDSVariables(M, LDSUsesInfo,
1197                                  KernelsThatIndirectlyAllocateDynamicLDS,
1198                                  DynamicVariables, OrderedKernels);
1199 
1200     // All kernel frames have been allocated. Calculate and record the
1201     // addresses.
1202 
1203     {
1204       const DataLayout &DL = M.getDataLayout();
1205 
1206       for (Function &Func : M.functions()) {
1207         if (Func.isDeclaration() || !isKernelLDS(&Func))
1208           continue;
1209 
1210         // All three of these are optional. The first variable is allocated at
1211         // zero. They are allocated by allocateKnownAddressLDSGlobal in the
1212         // following order:
1213         //{
1214         //  module.lds
1215         //  alignment padding
1216         //  kernel instance
1217         //  alignment padding
1218         //  dynamic lds variables
1219         //}
1220 
1221         const bool AllocateModuleScopeStruct =
1222             MaybeModuleScopeStruct && !canElideModuleLDS(Func);
1223 
1224         auto Replacement = KernelToReplacement.find(&Func);
1225         const bool AllocateKernelScopeStruct =
1226             Replacement != KernelToReplacement.end();
1227 
1228         const bool AllocateDynamicVariable =
1229             KernelToCreatedDynamicLDS.contains(&Func);
1230 
1231         uint32_t Offset = 0;
1232 
1233         if (AllocateModuleScopeStruct) {
1234           // Allocated at zero, recorded once on construction, not once per
1235           // kernel
1236           Offset += DL.getTypeAllocSize(MaybeModuleScopeStruct->getValueType());
1237         }
1238 
1239         if (AllocateKernelScopeStruct) {
1240           GlobalVariable *KernelStruct = Replacement->second.SGV;
1241 
1242           Offset = alignTo(Offset, AMDGPU::getAlign(DL, KernelStruct));
1243 
1244           recordLDSAbsoluteAddress(&M, KernelStruct, Offset);
1245 
1246           Offset += DL.getTypeAllocSize(KernelStruct->getValueType());
1247 
1248         }
1249 
1250         if (AllocateDynamicVariable) {
1251           GlobalVariable *DynamicVariable = KernelToCreatedDynamicLDS[&Func];
1252 
1253           Offset = alignTo(Offset, AMDGPU::getAlign(DL, DynamicVariable));
1254 
1255           recordLDSAbsoluteAddress(&M, DynamicVariable, Offset);
1256         }
1257       }
1258     }
1259 
1260     for (auto &GV : make_early_inc_range(M.globals()))
1261       if (AMDGPU::isLDSVariableToLower(GV)) {
1262         // probably want to remove from used lists
1263         GV.removeDeadConstantUsers();
1264         if (GV.use_empty())
1265           GV.eraseFromParent();
1266       }
1267 
1268     return Changed;
1269   }
1270 
1271 private:
1272   // Increase the alignment of LDS globals if necessary to maximise the chance
1273   // that we can use aligned LDS instructions to access them.
1274   static bool superAlignLDSGlobals(Module &M) {
1275     const DataLayout &DL = M.getDataLayout();
1276     bool Changed = false;
1277     if (!SuperAlignLDSGlobals) {
1278       return Changed;
1279     }
1280 
1281     for (auto &GV : M.globals()) {
1282       if (GV.getType()->getPointerAddressSpace() != AMDGPUAS::LOCAL_ADDRESS) {
1283         // Only changing alignment of LDS variables
1284         continue;
1285       }
1286       if (!GV.hasInitializer()) {
1287         // cuda/hip extern __shared__ variable, leave alignment alone
1288         continue;
1289       }
1290 
1291       Align Alignment = AMDGPU::getAlign(DL, &GV);
1292       TypeSize GVSize = DL.getTypeAllocSize(GV.getValueType());
1293 
1294       if (GVSize > 8) {
1295         // We might want to use a b96 or b128 load/store
1296         Alignment = std::max(Alignment, Align(16));
1297       } else if (GVSize > 4) {
1298         // We might want to use a b64 load/store
1299         Alignment = std::max(Alignment, Align(8));
1300       } else if (GVSize > 2) {
1301         // We might want to use a b32 load/store
1302         Alignment = std::max(Alignment, Align(4));
1303       } else if (GVSize > 1) {
1304         // We might want to use a b16 load/store
1305         Alignment = std::max(Alignment, Align(2));
1306       }
1307 
1308       if (Alignment != AMDGPU::getAlign(DL, &GV)) {
1309         Changed = true;
1310         GV.setAlignment(Alignment);
1311       }
1312     }
1313     return Changed;
1314   }
1315 
1316   static LDSVariableReplacement createLDSVariableReplacement(
1317       Module &M, std::string VarName,
1318       DenseSet<GlobalVariable *> const &LDSVarsToTransform) {
1319     // Create a struct instance containing LDSVarsToTransform and map from those
1320     // variables to ConstantExprGEP
1321     // Variables may be introduced to meet alignment requirements. No aliasing
1322     // metadata is useful for these as they have no uses. Erased before return.
1323 
1324     LLVMContext &Ctx = M.getContext();
1325     const DataLayout &DL = M.getDataLayout();
1326     assert(!LDSVarsToTransform.empty());
1327 
1328     SmallVector<OptimizedStructLayoutField, 8> LayoutFields;
1329     LayoutFields.reserve(LDSVarsToTransform.size());
1330     {
1331       // The order of fields in this struct depends on the order of
1332       // varables in the argument which varies when changing how they
1333       // are identified, leading to spurious test breakage.
1334       auto Sorted = sortByName(
1335           std::vector(LDSVarsToTransform.begin(), LDSVarsToTransform.end()));
1336 
1337       for (GlobalVariable *GV : Sorted) {
1338         OptimizedStructLayoutField F(GV,
1339                                      DL.getTypeAllocSize(GV->getValueType()),
1340                                      AMDGPU::getAlign(DL, GV));
1341         LayoutFields.emplace_back(F);
1342       }
1343     }
1344 
1345     performOptimizedStructLayout(LayoutFields);
1346 
1347     std::vector<GlobalVariable *> LocalVars;
1348     BitVector IsPaddingField;
1349     LocalVars.reserve(LDSVarsToTransform.size()); // will be at least this large
1350     IsPaddingField.reserve(LDSVarsToTransform.size());
1351     {
1352       uint64_t CurrentOffset = 0;
1353       for (size_t I = 0; I < LayoutFields.size(); I++) {
1354         GlobalVariable *FGV = static_cast<GlobalVariable *>(
1355             const_cast<void *>(LayoutFields[I].Id));
1356         Align DataAlign = LayoutFields[I].Alignment;
1357 
1358         uint64_t DataAlignV = DataAlign.value();
1359         if (uint64_t Rem = CurrentOffset % DataAlignV) {
1360           uint64_t Padding = DataAlignV - Rem;
1361 
1362           // Append an array of padding bytes to meet alignment requested
1363           // Note (o +      (a - (o % a)) ) % a == 0
1364           //      (offset + Padding       ) % align == 0
1365 
1366           Type *ATy = ArrayType::get(Type::getInt8Ty(Ctx), Padding);
1367           LocalVars.push_back(new GlobalVariable(
1368               M, ATy, false, GlobalValue::InternalLinkage, UndefValue::get(ATy),
1369               "", nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS,
1370               false));
1371           IsPaddingField.push_back(true);
1372           CurrentOffset += Padding;
1373         }
1374 
1375         LocalVars.push_back(FGV);
1376         IsPaddingField.push_back(false);
1377         CurrentOffset += LayoutFields[I].Size;
1378       }
1379     }
1380 
1381     std::vector<Type *> LocalVarTypes;
1382     LocalVarTypes.reserve(LocalVars.size());
1383     std::transform(
1384         LocalVars.cbegin(), LocalVars.cend(), std::back_inserter(LocalVarTypes),
1385         [](const GlobalVariable *V) -> Type * { return V->getValueType(); });
1386 
1387     StructType *LDSTy = StructType::create(Ctx, LocalVarTypes, VarName + ".t");
1388 
1389     Align StructAlign = AMDGPU::getAlign(DL, LocalVars[0]);
1390 
1391     GlobalVariable *SGV = new GlobalVariable(
1392         M, LDSTy, false, GlobalValue::InternalLinkage, UndefValue::get(LDSTy),
1393         VarName, nullptr, GlobalValue::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS,
1394         false);
1395     SGV->setAlignment(StructAlign);
1396 
1397     DenseMap<GlobalVariable *, Constant *> Map;
1398     Type *I32 = Type::getInt32Ty(Ctx);
1399     for (size_t I = 0; I < LocalVars.size(); I++) {
1400       GlobalVariable *GV = LocalVars[I];
1401       Constant *GEPIdx[] = {ConstantInt::get(I32, 0), ConstantInt::get(I32, I)};
1402       Constant *GEP = ConstantExpr::getGetElementPtr(LDSTy, SGV, GEPIdx, true);
1403       if (IsPaddingField[I]) {
1404         assert(GV->use_empty());
1405         GV->eraseFromParent();
1406       } else {
1407         Map[GV] = GEP;
1408       }
1409     }
1410     assert(Map.size() == LDSVarsToTransform.size());
1411     return {SGV, std::move(Map)};
1412   }
1413 
1414   template <typename PredicateTy>
1415   static void replaceLDSVariablesWithStruct(
1416       Module &M, DenseSet<GlobalVariable *> const &LDSVarsToTransformArg,
1417       const LDSVariableReplacement &Replacement, PredicateTy Predicate) {
1418     LLVMContext &Ctx = M.getContext();
1419     const DataLayout &DL = M.getDataLayout();
1420 
1421     // A hack... we need to insert the aliasing info in a predictable order for
1422     // lit tests. Would like to have them in a stable order already, ideally the
1423     // same order they get allocated, which might mean an ordered set container
1424     std::vector<GlobalVariable *> LDSVarsToTransform = sortByName(std::vector(
1425         LDSVarsToTransformArg.begin(), LDSVarsToTransformArg.end()));
1426 
1427     // Create alias.scope and their lists. Each field in the new structure
1428     // does not alias with all other fields.
1429     SmallVector<MDNode *> AliasScopes;
1430     SmallVector<Metadata *> NoAliasList;
1431     const size_t NumberVars = LDSVarsToTransform.size();
1432     if (NumberVars > 1) {
1433       MDBuilder MDB(Ctx);
1434       AliasScopes.reserve(NumberVars);
1435       MDNode *Domain = MDB.createAnonymousAliasScopeDomain();
1436       for (size_t I = 0; I < NumberVars; I++) {
1437         MDNode *Scope = MDB.createAnonymousAliasScope(Domain);
1438         AliasScopes.push_back(Scope);
1439       }
1440       NoAliasList.append(&AliasScopes[1], AliasScopes.end());
1441     }
1442 
1443     // Replace uses of ith variable with a constantexpr to the corresponding
1444     // field of the instance that will be allocated by AMDGPUMachineFunction
1445     for (size_t I = 0; I < NumberVars; I++) {
1446       GlobalVariable *GV = LDSVarsToTransform[I];
1447       Constant *GEP = Replacement.LDSVarsToConstantGEP[GV];
1448 
1449       GV->replaceUsesWithIf(GEP, Predicate);
1450 
1451       APInt APOff(DL.getIndexTypeSizeInBits(GEP->getType()), 0);
1452       GEP->stripAndAccumulateInBoundsConstantOffsets(DL, APOff);
1453       uint64_t Offset = APOff.getZExtValue();
1454 
1455       Align A =
1456           commonAlignment(Replacement.SGV->getAlign().valueOrOne(), Offset);
1457 
1458       if (I)
1459         NoAliasList[I - 1] = AliasScopes[I - 1];
1460       MDNode *NoAlias =
1461           NoAliasList.empty() ? nullptr : MDNode::get(Ctx, NoAliasList);
1462       MDNode *AliasScope =
1463           AliasScopes.empty() ? nullptr : MDNode::get(Ctx, {AliasScopes[I]});
1464 
1465       refineUsesAlignmentAndAA(GEP, A, DL, AliasScope, NoAlias);
1466     }
1467   }
1468 
1469   static void refineUsesAlignmentAndAA(Value *Ptr, Align A,
1470                                        const DataLayout &DL, MDNode *AliasScope,
1471                                        MDNode *NoAlias, unsigned MaxDepth = 5) {
1472     if (!MaxDepth || (A == 1 && !AliasScope))
1473       return;
1474 
1475     for (User *U : Ptr->users()) {
1476       if (auto *I = dyn_cast<Instruction>(U)) {
1477         if (AliasScope && I->mayReadOrWriteMemory()) {
1478           MDNode *AS = I->getMetadata(LLVMContext::MD_alias_scope);
1479           AS = (AS ? MDNode::getMostGenericAliasScope(AS, AliasScope)
1480                    : AliasScope);
1481           I->setMetadata(LLVMContext::MD_alias_scope, AS);
1482 
1483           MDNode *NA = I->getMetadata(LLVMContext::MD_noalias);
1484           NA = (NA ? MDNode::intersect(NA, NoAlias) : NoAlias);
1485           I->setMetadata(LLVMContext::MD_noalias, NA);
1486         }
1487       }
1488 
1489       if (auto *LI = dyn_cast<LoadInst>(U)) {
1490         LI->setAlignment(std::max(A, LI->getAlign()));
1491         continue;
1492       }
1493       if (auto *SI = dyn_cast<StoreInst>(U)) {
1494         if (SI->getPointerOperand() == Ptr)
1495           SI->setAlignment(std::max(A, SI->getAlign()));
1496         continue;
1497       }
1498       if (auto *AI = dyn_cast<AtomicRMWInst>(U)) {
1499         // None of atomicrmw operations can work on pointers, but let's
1500         // check it anyway in case it will or we will process ConstantExpr.
1501         if (AI->getPointerOperand() == Ptr)
1502           AI->setAlignment(std::max(A, AI->getAlign()));
1503         continue;
1504       }
1505       if (auto *AI = dyn_cast<AtomicCmpXchgInst>(U)) {
1506         if (AI->getPointerOperand() == Ptr)
1507           AI->setAlignment(std::max(A, AI->getAlign()));
1508         continue;
1509       }
1510       if (auto *GEP = dyn_cast<GetElementPtrInst>(U)) {
1511         unsigned BitWidth = DL.getIndexTypeSizeInBits(GEP->getType());
1512         APInt Off(BitWidth, 0);
1513         if (GEP->getPointerOperand() == Ptr) {
1514           Align GA;
1515           if (GEP->accumulateConstantOffset(DL, Off))
1516             GA = commonAlignment(A, Off.getLimitedValue());
1517           refineUsesAlignmentAndAA(GEP, GA, DL, AliasScope, NoAlias,
1518                                    MaxDepth - 1);
1519         }
1520         continue;
1521       }
1522       if (auto *I = dyn_cast<Instruction>(U)) {
1523         if (I->getOpcode() == Instruction::BitCast ||
1524             I->getOpcode() == Instruction::AddrSpaceCast)
1525           refineUsesAlignmentAndAA(I, A, DL, AliasScope, NoAlias, MaxDepth - 1);
1526       }
1527     }
1528   }
1529 };
1530 
1531 } // namespace
1532 char AMDGPULowerModuleLDS::ID = 0;
1533 
1534 char &llvm::AMDGPULowerModuleLDSID = AMDGPULowerModuleLDS::ID;
1535 
1536 INITIALIZE_PASS(AMDGPULowerModuleLDS, DEBUG_TYPE,
1537                 "Lower uses of LDS variables from non-kernel functions", false,
1538                 false)
1539 
1540 ModulePass *llvm::createAMDGPULowerModuleLDSPass() {
1541   return new AMDGPULowerModuleLDS();
1542 }
1543 
1544 PreservedAnalyses AMDGPULowerModuleLDSPass::run(Module &M,
1545                                                 ModuleAnalysisManager &) {
1546   return AMDGPULowerModuleLDS().runOnModule(M) ? PreservedAnalyses::none()
1547                                                : PreservedAnalyses::all();
1548 }
1549