1 | //===- PreISelIntrinsicLowering.cpp - Pre-ISel intrinsic lowering pass ----===// |
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 implements IR lowering for the llvm.memcpy, llvm.memmove, |
10 | // llvm.memset, llvm.load.relative and llvm.objc.* intrinsics. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "llvm/CodeGen/PreISelIntrinsicLowering.h" |
15 | #include "llvm/Analysis/ObjCARCInstKind.h" |
16 | #include "llvm/Analysis/ObjCARCUtil.h" |
17 | #include "llvm/Analysis/TargetLibraryInfo.h" |
18 | #include "llvm/Analysis/TargetTransformInfo.h" |
19 | #include "llvm/CodeGen/ExpandVectorPredication.h" |
20 | #include "llvm/CodeGen/Passes.h" |
21 | #include "llvm/CodeGen/TargetLowering.h" |
22 | #include "llvm/CodeGen/TargetPassConfig.h" |
23 | #include "llvm/IR/Function.h" |
24 | #include "llvm/IR/IRBuilder.h" |
25 | #include "llvm/IR/Instructions.h" |
26 | #include "llvm/IR/IntrinsicInst.h" |
27 | #include "llvm/IR/Module.h" |
28 | #include "llvm/IR/Type.h" |
29 | #include "llvm/IR/Use.h" |
30 | #include "llvm/InitializePasses.h" |
31 | #include "llvm/Pass.h" |
32 | #include "llvm/Support/Casting.h" |
33 | #include "llvm/Target/TargetMachine.h" |
34 | #include "llvm/Transforms/Scalar/LowerConstantIntrinsics.h" |
35 | #include "llvm/Transforms/Utils/BuildLibCalls.h" |
36 | #include "llvm/Transforms/Utils/LowerMemIntrinsics.h" |
37 | #include "llvm/Transforms/Utils/LowerVectorIntrinsics.h" |
38 | |
39 | using namespace llvm; |
40 | |
41 | /// Threshold to leave statically sized memory intrinsic calls. Calls of known |
42 | /// size larger than this will be expanded by the pass. Calls of unknown or |
43 | /// lower size will be left for expansion in codegen. |
44 | static cl::opt<int64_t> MemIntrinsicExpandSizeThresholdOpt( |
45 | "mem-intrinsic-expand-size" , |
46 | cl::desc("Set minimum mem intrinsic size to expand in IR" ), cl::init(Val: -1), |
47 | cl::Hidden); |
48 | |
49 | namespace { |
50 | |
51 | struct PreISelIntrinsicLowering { |
52 | const TargetMachine *TM; |
53 | const function_ref<TargetTransformInfo &(Function &)> LookupTTI; |
54 | const function_ref<TargetLibraryInfo &(Function &)> LookupTLI; |
55 | |
56 | /// If this is true, assume it's preferably to leave memory intrinsic calls |
57 | /// for replacement with a library call later. Otherwise this depends on |
58 | /// TargetLoweringInfo availability of the corresponding function. |
59 | const bool UseMemIntrinsicLibFunc; |
60 | |
61 | explicit PreISelIntrinsicLowering( |
62 | const TargetMachine *TM_, |
63 | function_ref<TargetTransformInfo &(Function &)> LookupTTI_, |
64 | function_ref<TargetLibraryInfo &(Function &)> LookupTLI_, |
65 | bool UseMemIntrinsicLibFunc_ = true) |
66 | : TM(TM_), LookupTTI(LookupTTI_), LookupTLI(LookupTLI_), |
67 | UseMemIntrinsicLibFunc(UseMemIntrinsicLibFunc_) {} |
68 | |
69 | static bool shouldExpandMemIntrinsicWithSize(Value *Size, |
70 | const TargetTransformInfo &TTI); |
71 | bool |
72 | expandMemIntrinsicUses(Function &F, |
73 | DenseMap<Constant *, GlobalVariable *> &CMap) const; |
74 | bool lowerIntrinsics(Module &M) const; |
75 | }; |
76 | |
77 | } // namespace |
78 | |
79 | template <class T> static bool forEachCall(Function &Intrin, T Callback) { |
80 | // Lowering all intrinsics in a function will delete multiple uses, so we |
81 | // can't use an early-inc-range. In case some remain, we don't want to look |
82 | // at them again. Unfortunately, Value::UseList is private, so we can't use a |
83 | // simple Use**. If LastUse is null, the next use to consider is |
84 | // Intrin.use_begin(), otherwise it's LastUse->getNext(). |
85 | Use *LastUse = nullptr; |
86 | bool Changed = false; |
87 | while (!Intrin.use_empty() && (!LastUse || LastUse->getNext())) { |
88 | Use *U = LastUse ? LastUse->getNext() : &*Intrin.use_begin(); |
89 | bool Removed = false; |
90 | // An intrinsic cannot have its address taken, so it cannot be an argument |
91 | // operand. It might be used as operand in debug metadata, though. |
92 | if (auto CI = dyn_cast<CallInst>(Val: U->getUser())) |
93 | Changed |= Removed = Callback(CI); |
94 | if (!Removed) |
95 | LastUse = U; |
96 | } |
97 | return Changed; |
98 | } |
99 | |
100 | static bool lowerLoadRelative(Function &F) { |
101 | if (F.use_empty()) |
102 | return false; |
103 | |
104 | bool Changed = false; |
105 | Type *Int32Ty = Type::getInt32Ty(C&: F.getContext()); |
106 | |
107 | for (Use &U : llvm::make_early_inc_range(Range: F.uses())) { |
108 | auto CI = dyn_cast<CallInst>(Val: U.getUser()); |
109 | if (!CI || CI->getCalledOperand() != &F) |
110 | continue; |
111 | |
112 | IRBuilder<> B(CI); |
113 | Value *OffsetPtr = |
114 | B.CreatePtrAdd(Ptr: CI->getArgOperand(i: 0), Offset: CI->getArgOperand(i: 1)); |
115 | Value *OffsetI32 = B.CreateAlignedLoad(Ty: Int32Ty, Ptr: OffsetPtr, Align: Align(4)); |
116 | |
117 | Value *ResultPtr = B.CreatePtrAdd(Ptr: CI->getArgOperand(i: 0), Offset: OffsetI32); |
118 | |
119 | CI->replaceAllUsesWith(V: ResultPtr); |
120 | CI->eraseFromParent(); |
121 | Changed = true; |
122 | } |
123 | |
124 | return Changed; |
125 | } |
126 | |
127 | // ObjCARC has knowledge about whether an obj-c runtime function needs to be |
128 | // always tail-called or never tail-called. |
129 | static CallInst::TailCallKind getOverridingTailCallKind(const Function &F) { |
130 | objcarc::ARCInstKind Kind = objcarc::GetFunctionClass(F: &F); |
131 | if (objcarc::IsAlwaysTail(Class: Kind)) |
132 | return CallInst::TCK_Tail; |
133 | else if (objcarc::IsNeverTail(Class: Kind)) |
134 | return CallInst::TCK_NoTail; |
135 | return CallInst::TCK_None; |
136 | } |
137 | |
138 | static bool lowerObjCCall(Function &F, const char *NewFn, |
139 | bool setNonLazyBind = false) { |
140 | assert(IntrinsicInst::mayLowerToFunctionCall(F.getIntrinsicID()) && |
141 | "Pre-ISel intrinsics do lower into regular function calls" ); |
142 | if (F.use_empty()) |
143 | return false; |
144 | |
145 | // If we haven't already looked up this function, check to see if the |
146 | // program already contains a function with this name. |
147 | Module *M = F.getParent(); |
148 | FunctionCallee FCache = M->getOrInsertFunction(Name: NewFn, T: F.getFunctionType()); |
149 | |
150 | if (Function *Fn = dyn_cast<Function>(Val: FCache.getCallee())) { |
151 | Fn->setLinkage(F.getLinkage()); |
152 | if (setNonLazyBind && !Fn->isWeakForLinker()) { |
153 | // If we have Native ARC, set nonlazybind attribute for these APIs for |
154 | // performance. |
155 | Fn->addFnAttr(Kind: Attribute::NonLazyBind); |
156 | } |
157 | } |
158 | |
159 | CallInst::TailCallKind OverridingTCK = getOverridingTailCallKind(F); |
160 | |
161 | for (Use &U : llvm::make_early_inc_range(Range: F.uses())) { |
162 | auto *CB = cast<CallBase>(Val: U.getUser()); |
163 | |
164 | if (CB->getCalledFunction() != &F) { |
165 | assert(objcarc::getAttachedARCFunction(CB) == &F && |
166 | "use expected to be the argument of operand bundle " |
167 | "\"clang.arc.attachedcall\"" ); |
168 | U.set(FCache.getCallee()); |
169 | continue; |
170 | } |
171 | |
172 | auto *CI = cast<CallInst>(Val: CB); |
173 | assert(CI->getCalledFunction() && "Cannot lower an indirect call!" ); |
174 | |
175 | IRBuilder<> Builder(CI->getParent(), CI->getIterator()); |
176 | SmallVector<Value *, 8> Args(CI->args()); |
177 | SmallVector<llvm::OperandBundleDef, 1> BundleList; |
178 | CI->getOperandBundlesAsDefs(Defs&: BundleList); |
179 | CallInst *NewCI = Builder.CreateCall(Callee: FCache, Args, OpBundles: BundleList); |
180 | NewCI->setName(CI->getName()); |
181 | |
182 | // Try to set the most appropriate TailCallKind based on both the current |
183 | // attributes and the ones that we could get from ObjCARC's special |
184 | // knowledge of the runtime functions. |
185 | // |
186 | // std::max respects both requirements of notail and tail here: |
187 | // * notail on either the call or from ObjCARC becomes notail |
188 | // * tail on either side is stronger than none, but not notail |
189 | CallInst::TailCallKind TCK = CI->getTailCallKind(); |
190 | NewCI->setTailCallKind(std::max(a: TCK, b: OverridingTCK)); |
191 | |
192 | // Transfer the 'returned' attribute from the intrinsic to the call site. |
193 | // By applying this only to intrinsic call sites, we avoid applying it to |
194 | // non-ARC explicit calls to things like objc_retain which have not been |
195 | // auto-upgraded to use the intrinsics. |
196 | unsigned Index; |
197 | if (F.getAttributes().hasAttrSomewhere(Kind: Attribute::Returned, Index: &Index) && |
198 | Index) |
199 | NewCI->addParamAttr(ArgNo: Index - AttributeList::FirstArgIndex, |
200 | Kind: Attribute::Returned); |
201 | |
202 | if (!CI->use_empty()) |
203 | CI->replaceAllUsesWith(V: NewCI); |
204 | CI->eraseFromParent(); |
205 | } |
206 | |
207 | return true; |
208 | } |
209 | |
210 | // TODO: Should refine based on estimated number of accesses (e.g. does it |
211 | // require splitting based on alignment) |
212 | bool PreISelIntrinsicLowering::shouldExpandMemIntrinsicWithSize( |
213 | Value *Size, const TargetTransformInfo &TTI) { |
214 | ConstantInt *CI = dyn_cast<ConstantInt>(Val: Size); |
215 | if (!CI) |
216 | return true; |
217 | uint64_t Threshold = MemIntrinsicExpandSizeThresholdOpt.getNumOccurrences() |
218 | ? MemIntrinsicExpandSizeThresholdOpt |
219 | : TTI.getMaxMemIntrinsicInlineSizeThreshold(); |
220 | uint64_t SizeVal = CI->getZExtValue(); |
221 | |
222 | // Treat a threshold of 0 as a special case to force expansion of all |
223 | // intrinsics, including size 0. |
224 | return SizeVal > Threshold || Threshold == 0; |
225 | } |
226 | |
227 | static bool canEmitLibcall(const TargetMachine *TM, Function *F, |
228 | RTLIB::Libcall LC) { |
229 | // TODO: Should this consider the address space of the memcpy? |
230 | if (!TM) |
231 | return true; |
232 | const TargetLowering *TLI = TM->getSubtargetImpl(*F)->getTargetLowering(); |
233 | return TLI->getLibcallName(Call: LC) != nullptr; |
234 | } |
235 | |
236 | static bool canEmitMemcpy(const TargetMachine *TM, Function *F) { |
237 | // TODO: Should this consider the address space of the memcpy? |
238 | if (!TM) |
239 | return true; |
240 | const TargetLowering *TLI = TM->getSubtargetImpl(*F)->getTargetLowering(); |
241 | return TLI->getMemcpyName() != nullptr; |
242 | } |
243 | |
244 | // Return a value appropriate for use with the memset_pattern16 libcall, if |
245 | // possible and if we know how. (Adapted from equivalent helper in |
246 | // LoopIdiomRecognize). |
247 | static Constant *getMemSetPattern16Value(MemSetPatternInst *Inst, |
248 | const TargetLibraryInfo &TLI) { |
249 | // TODO: This could check for UndefValue because it can be merged into any |
250 | // other valid pattern. |
251 | |
252 | // Don't emit libcalls if a non-default address space is being used. |
253 | if (Inst->getRawDest()->getType()->getPointerAddressSpace() != 0) |
254 | return nullptr; |
255 | |
256 | Value *V = Inst->getValue(); |
257 | Type *VTy = V->getType(); |
258 | const DataLayout &DL = Inst->getDataLayout(); |
259 | Module *M = Inst->getModule(); |
260 | |
261 | if (!isLibFuncEmittable(M, TLI: &TLI, TheLibFunc: LibFunc_memset_pattern16)) |
262 | return nullptr; |
263 | |
264 | // If the value isn't a constant, we can't promote it to being in a constant |
265 | // array. We could theoretically do a store to an alloca or something, but |
266 | // that doesn't seem worthwhile. |
267 | Constant *C = dyn_cast<Constant>(Val: V); |
268 | if (!C || isa<ConstantExpr>(Val: C)) |
269 | return nullptr; |
270 | |
271 | // Only handle simple values that are a power of two bytes in size. |
272 | uint64_t Size = DL.getTypeSizeInBits(Ty: VTy); |
273 | if (!DL.typeSizeEqualsStoreSize(Ty: VTy) || !isPowerOf2_64(Value: Size)) |
274 | return nullptr; |
275 | |
276 | // Don't care enough about darwin/ppc to implement this. |
277 | if (DL.isBigEndian()) |
278 | return nullptr; |
279 | |
280 | // Convert to size in bytes. |
281 | Size /= 8; |
282 | |
283 | // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see |
284 | // if the top and bottom are the same (e.g. for vectors and large integers). |
285 | if (Size > 16) |
286 | return nullptr; |
287 | |
288 | // If the constant is exactly 16 bytes, just use it. |
289 | if (Size == 16) |
290 | return C; |
291 | |
292 | // Otherwise, we'll use an array of the constants. |
293 | uint64_t ArraySize = 16 / Size; |
294 | ArrayType *AT = ArrayType::get(ElementType: V->getType(), NumElements: ArraySize); |
295 | return ConstantArray::get(T: AT, V: std::vector<Constant *>(ArraySize, C)); |
296 | } |
297 | |
298 | // TODO: Handle atomic memcpy and memcpy.inline |
299 | // TODO: Pass ScalarEvolution |
300 | bool PreISelIntrinsicLowering::expandMemIntrinsicUses( |
301 | Function &F, DenseMap<Constant *, GlobalVariable *> &CMap) const { |
302 | Intrinsic::ID ID = F.getIntrinsicID(); |
303 | bool Changed = false; |
304 | |
305 | for (User *U : llvm::make_early_inc_range(Range: F.users())) { |
306 | Instruction *Inst = cast<Instruction>(Val: U); |
307 | |
308 | switch (ID) { |
309 | case Intrinsic::memcpy: { |
310 | auto *Memcpy = cast<MemCpyInst>(Val: Inst); |
311 | Function *ParentFunc = Memcpy->getFunction(); |
312 | const TargetTransformInfo &TTI = LookupTTI(*ParentFunc); |
313 | if (shouldExpandMemIntrinsicWithSize(Size: Memcpy->getLength(), TTI)) { |
314 | if (UseMemIntrinsicLibFunc && canEmitMemcpy(TM, F: ParentFunc)) |
315 | break; |
316 | |
317 | // TODO: For optsize, emit the loop into a separate function |
318 | expandMemCpyAsLoop(MemCpy: Memcpy, TTI); |
319 | Changed = true; |
320 | Memcpy->eraseFromParent(); |
321 | } |
322 | |
323 | break; |
324 | } |
325 | case Intrinsic::memcpy_inline: { |
326 | // Only expand llvm.memcpy.inline with non-constant length in this |
327 | // codepath, leaving the current SelectionDAG expansion for constant |
328 | // length memcpy intrinsics undisturbed. |
329 | auto *Memcpy = cast<MemCpyInst>(Val: Inst); |
330 | if (isa<ConstantInt>(Val: Memcpy->getLength())) |
331 | break; |
332 | |
333 | Function *ParentFunc = Memcpy->getFunction(); |
334 | const TargetTransformInfo &TTI = LookupTTI(*ParentFunc); |
335 | expandMemCpyAsLoop(MemCpy: Memcpy, TTI); |
336 | Changed = true; |
337 | Memcpy->eraseFromParent(); |
338 | break; |
339 | } |
340 | case Intrinsic::memmove: { |
341 | auto *Memmove = cast<MemMoveInst>(Val: Inst); |
342 | Function *ParentFunc = Memmove->getFunction(); |
343 | const TargetTransformInfo &TTI = LookupTTI(*ParentFunc); |
344 | if (shouldExpandMemIntrinsicWithSize(Size: Memmove->getLength(), TTI)) { |
345 | if (UseMemIntrinsicLibFunc && |
346 | canEmitLibcall(TM, F: ParentFunc, LC: RTLIB::MEMMOVE)) |
347 | break; |
348 | |
349 | if (expandMemMoveAsLoop(MemMove: Memmove, TTI)) { |
350 | Changed = true; |
351 | Memmove->eraseFromParent(); |
352 | } |
353 | } |
354 | |
355 | break; |
356 | } |
357 | case Intrinsic::memset: { |
358 | auto *Memset = cast<MemSetInst>(Val: Inst); |
359 | Function *ParentFunc = Memset->getFunction(); |
360 | const TargetTransformInfo &TTI = LookupTTI(*ParentFunc); |
361 | if (shouldExpandMemIntrinsicWithSize(Size: Memset->getLength(), TTI)) { |
362 | if (UseMemIntrinsicLibFunc && |
363 | canEmitLibcall(TM, F: ParentFunc, LC: RTLIB::MEMSET)) |
364 | break; |
365 | |
366 | expandMemSetAsLoop(MemSet: Memset); |
367 | Changed = true; |
368 | Memset->eraseFromParent(); |
369 | } |
370 | |
371 | break; |
372 | } |
373 | case Intrinsic::memset_inline: { |
374 | // Only expand llvm.memset.inline with non-constant length in this |
375 | // codepath, leaving the current SelectionDAG expansion for constant |
376 | // length memset intrinsics undisturbed. |
377 | auto *Memset = cast<MemSetInst>(Val: Inst); |
378 | if (isa<ConstantInt>(Val: Memset->getLength())) |
379 | break; |
380 | |
381 | expandMemSetAsLoop(MemSet: Memset); |
382 | Changed = true; |
383 | Memset->eraseFromParent(); |
384 | break; |
385 | } |
386 | case Intrinsic::experimental_memset_pattern: { |
387 | auto *Memset = cast<MemSetPatternInst>(Val: Inst); |
388 | const TargetLibraryInfo &TLI = LookupTLI(*Memset->getFunction()); |
389 | Constant *PatternValue = getMemSetPattern16Value(Inst: Memset, TLI); |
390 | if (!PatternValue) { |
391 | // If it isn't possible to emit a memset_pattern16 libcall, expand to |
392 | // a loop instead. |
393 | expandMemSetPatternAsLoop(MemSet: Memset); |
394 | Changed = true; |
395 | Memset->eraseFromParent(); |
396 | break; |
397 | } |
398 | // FIXME: There is currently no profitability calculation for emitting |
399 | // the libcall vs expanding the memset.pattern directly. |
400 | IRBuilder<> Builder(Inst); |
401 | Module *M = Memset->getModule(); |
402 | const DataLayout &DL = Memset->getDataLayout(); |
403 | |
404 | Type *DestPtrTy = Memset->getRawDest()->getType(); |
405 | Type *SizeTTy = TLI.getSizeTType(M: *M); |
406 | StringRef FuncName = "memset_pattern16" ; |
407 | FunctionCallee MSP = getOrInsertLibFunc(M, TLI, TheLibFunc: LibFunc_memset_pattern16, |
408 | RetTy: Builder.getVoidTy(), Args: DestPtrTy, |
409 | Args: Builder.getPtrTy(), Args: SizeTTy); |
410 | inferNonMandatoryLibFuncAttrs(M, Name: FuncName, TLI); |
411 | |
412 | // Otherwise we should form a memset_pattern16. PatternValue is known |
413 | // to be an constant array of 16-bytes. Put the value into a mergable |
414 | // global. |
415 | assert(Memset->getRawDest()->getType()->getPointerAddressSpace() == 0 && |
416 | "Should have skipped if non-zero AS" ); |
417 | GlobalVariable *GV; |
418 | auto It = CMap.find(Val: PatternValue); |
419 | if (It != CMap.end()) { |
420 | GV = It->second; |
421 | } else { |
422 | GV = new GlobalVariable( |
423 | *M, PatternValue->getType(), /*isConstant=*/true, |
424 | GlobalValue::PrivateLinkage, PatternValue, ".memset_pattern" ); |
425 | GV->setUnnamedAddr( |
426 | GlobalValue::UnnamedAddr::Global); // Ok to merge these. |
427 | // TODO: Consider relaxing alignment requirement. |
428 | GV->setAlignment(Align(16)); |
429 | CMap[PatternValue] = GV; |
430 | } |
431 | Value *PatternPtr = GV; |
432 | Value *NumBytes = Builder.CreateMul( |
433 | LHS: TLI.getAsSizeT(V: DL.getTypeAllocSize(Ty: Memset->getValue()->getType()), |
434 | M: *M), |
435 | RHS: Builder.CreateZExtOrTrunc(V: Memset->getLength(), DestTy: SizeTTy)); |
436 | CallInst *MemsetPattern16Call = |
437 | Builder.CreateCall(Callee: MSP, Args: {Memset->getRawDest(), PatternPtr, NumBytes}); |
438 | MemsetPattern16Call->setAAMetadata(Memset->getAAMetadata()); |
439 | // Preserve any call site attributes on the destination pointer |
440 | // argument (e.g. alignment). |
441 | AttrBuilder ArgAttrs(Memset->getContext(), |
442 | Memset->getAttributes().getParamAttrs(ArgNo: 0)); |
443 | MemsetPattern16Call->setAttributes( |
444 | MemsetPattern16Call->getAttributes().addParamAttributes( |
445 | C&: Memset->getContext(), ArgNo: 0, B: ArgAttrs)); |
446 | Changed = true; |
447 | Memset->eraseFromParent(); |
448 | break; |
449 | } |
450 | default: |
451 | llvm_unreachable("unhandled intrinsic" ); |
452 | } |
453 | } |
454 | |
455 | return Changed; |
456 | } |
457 | |
458 | bool PreISelIntrinsicLowering::lowerIntrinsics(Module &M) const { |
459 | // Map unique constants to globals. |
460 | DenseMap<Constant *, GlobalVariable *> CMap; |
461 | bool Changed = false; |
462 | for (Function &F : M) { |
463 | switch (F.getIntrinsicID()) { |
464 | default: |
465 | break; |
466 | case Intrinsic::memcpy: |
467 | case Intrinsic::memcpy_inline: |
468 | case Intrinsic::memmove: |
469 | case Intrinsic::memset: |
470 | case Intrinsic::memset_inline: |
471 | case Intrinsic::experimental_memset_pattern: |
472 | Changed |= expandMemIntrinsicUses(F, CMap); |
473 | break; |
474 | case Intrinsic::load_relative: |
475 | Changed |= lowerLoadRelative(F); |
476 | break; |
477 | case Intrinsic::is_constant: |
478 | case Intrinsic::objectsize: |
479 | Changed |= forEachCall(Intrin&: F, Callback: [&](CallInst *CI) { |
480 | Function *Parent = CI->getParent()->getParent(); |
481 | TargetLibraryInfo &TLI = LookupTLI(*Parent); |
482 | // Intrinsics in unreachable code are not lowered. |
483 | bool Changed = lowerConstantIntrinsics(F&: *Parent, TLI, /*DT=*/nullptr); |
484 | return Changed; |
485 | }); |
486 | break; |
487 | #define BEGIN_REGISTER_VP_INTRINSIC(VPID, MASKPOS, VLENPOS) \ |
488 | case Intrinsic::VPID: |
489 | #include "llvm/IR/VPIntrinsics.def" |
490 | forEachCall(Intrin&: F, Callback: [&](CallInst *CI) { |
491 | Function *Parent = CI->getParent()->getParent(); |
492 | const TargetTransformInfo &TTI = LookupTTI(*Parent); |
493 | auto *VPI = cast<VPIntrinsic>(Val: CI); |
494 | VPExpansionDetails ED = expandVectorPredicationIntrinsic(VPI&: *VPI, TTI); |
495 | // Expansion of VP intrinsics may change the IR but not actually |
496 | // replace the intrinsic, so update Changed for the pass |
497 | // and compute Removed for forEachCall. |
498 | Changed |= ED != VPExpansionDetails::IntrinsicUnchanged; |
499 | bool Removed = ED == VPExpansionDetails::IntrinsicReplaced; |
500 | return Removed; |
501 | }); |
502 | break; |
503 | case Intrinsic::objc_autorelease: |
504 | Changed |= lowerObjCCall(F, NewFn: "objc_autorelease" ); |
505 | break; |
506 | case Intrinsic::objc_autoreleasePoolPop: |
507 | Changed |= lowerObjCCall(F, NewFn: "objc_autoreleasePoolPop" ); |
508 | break; |
509 | case Intrinsic::objc_autoreleasePoolPush: |
510 | Changed |= lowerObjCCall(F, NewFn: "objc_autoreleasePoolPush" ); |
511 | break; |
512 | case Intrinsic::objc_autoreleaseReturnValue: |
513 | Changed |= lowerObjCCall(F, NewFn: "objc_autoreleaseReturnValue" ); |
514 | break; |
515 | case Intrinsic::objc_copyWeak: |
516 | Changed |= lowerObjCCall(F, NewFn: "objc_copyWeak" ); |
517 | break; |
518 | case Intrinsic::objc_destroyWeak: |
519 | Changed |= lowerObjCCall(F, NewFn: "objc_destroyWeak" ); |
520 | break; |
521 | case Intrinsic::objc_initWeak: |
522 | Changed |= lowerObjCCall(F, NewFn: "objc_initWeak" ); |
523 | break; |
524 | case Intrinsic::objc_loadWeak: |
525 | Changed |= lowerObjCCall(F, NewFn: "objc_loadWeak" ); |
526 | break; |
527 | case Intrinsic::objc_loadWeakRetained: |
528 | Changed |= lowerObjCCall(F, NewFn: "objc_loadWeakRetained" ); |
529 | break; |
530 | case Intrinsic::objc_moveWeak: |
531 | Changed |= lowerObjCCall(F, NewFn: "objc_moveWeak" ); |
532 | break; |
533 | case Intrinsic::objc_release: |
534 | Changed |= lowerObjCCall(F, NewFn: "objc_release" , setNonLazyBind: true); |
535 | break; |
536 | case Intrinsic::objc_retain: |
537 | Changed |= lowerObjCCall(F, NewFn: "objc_retain" , setNonLazyBind: true); |
538 | break; |
539 | case Intrinsic::objc_retainAutorelease: |
540 | Changed |= lowerObjCCall(F, NewFn: "objc_retainAutorelease" ); |
541 | break; |
542 | case Intrinsic::objc_retainAutoreleaseReturnValue: |
543 | Changed |= lowerObjCCall(F, NewFn: "objc_retainAutoreleaseReturnValue" ); |
544 | break; |
545 | case Intrinsic::objc_retainAutoreleasedReturnValue: |
546 | Changed |= lowerObjCCall(F, NewFn: "objc_retainAutoreleasedReturnValue" ); |
547 | break; |
548 | case Intrinsic::objc_claimAutoreleasedReturnValue: |
549 | Changed |= lowerObjCCall(F, NewFn: "objc_claimAutoreleasedReturnValue" ); |
550 | break; |
551 | case Intrinsic::objc_retainBlock: |
552 | Changed |= lowerObjCCall(F, NewFn: "objc_retainBlock" ); |
553 | break; |
554 | case Intrinsic::objc_storeStrong: |
555 | Changed |= lowerObjCCall(F, NewFn: "objc_storeStrong" ); |
556 | break; |
557 | case Intrinsic::objc_storeWeak: |
558 | Changed |= lowerObjCCall(F, NewFn: "objc_storeWeak" ); |
559 | break; |
560 | case Intrinsic::objc_unsafeClaimAutoreleasedReturnValue: |
561 | Changed |= lowerObjCCall(F, NewFn: "objc_unsafeClaimAutoreleasedReturnValue" ); |
562 | break; |
563 | case Intrinsic::objc_retainedObject: |
564 | Changed |= lowerObjCCall(F, NewFn: "objc_retainedObject" ); |
565 | break; |
566 | case Intrinsic::objc_unretainedObject: |
567 | Changed |= lowerObjCCall(F, NewFn: "objc_unretainedObject" ); |
568 | break; |
569 | case Intrinsic::objc_unretainedPointer: |
570 | Changed |= lowerObjCCall(F, NewFn: "objc_unretainedPointer" ); |
571 | break; |
572 | case Intrinsic::objc_retain_autorelease: |
573 | Changed |= lowerObjCCall(F, NewFn: "objc_retain_autorelease" ); |
574 | break; |
575 | case Intrinsic::objc_sync_enter: |
576 | Changed |= lowerObjCCall(F, NewFn: "objc_sync_enter" ); |
577 | break; |
578 | case Intrinsic::objc_sync_exit: |
579 | Changed |= lowerObjCCall(F, NewFn: "objc_sync_exit" ); |
580 | break; |
581 | case Intrinsic::exp: |
582 | case Intrinsic::exp2: |
583 | Changed |= forEachCall(Intrin&: F, Callback: [&](CallInst *CI) { |
584 | Type *Ty = CI->getArgOperand(i: 0)->getType(); |
585 | if (!isa<ScalableVectorType>(Val: Ty)) |
586 | return false; |
587 | const TargetLowering *TL = TM->getSubtargetImpl(F)->getTargetLowering(); |
588 | unsigned Op = TL->IntrinsicIDToISD(ID: F.getIntrinsicID()); |
589 | if (!TL->isOperationExpand(Op, VT: EVT::getEVT(Ty))) |
590 | return false; |
591 | return lowerUnaryVectorIntrinsicAsLoop(M, CI); |
592 | }); |
593 | break; |
594 | } |
595 | } |
596 | return Changed; |
597 | } |
598 | |
599 | namespace { |
600 | |
601 | class PreISelIntrinsicLoweringLegacyPass : public ModulePass { |
602 | public: |
603 | static char ID; |
604 | |
605 | PreISelIntrinsicLoweringLegacyPass() : ModulePass(ID) {} |
606 | |
607 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
608 | AU.addRequired<TargetTransformInfoWrapperPass>(); |
609 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
610 | AU.addRequired<TargetPassConfig>(); |
611 | } |
612 | |
613 | bool runOnModule(Module &M) override { |
614 | auto LookupTTI = [this](Function &F) -> TargetTransformInfo & { |
615 | return this->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); |
616 | }; |
617 | auto LookupTLI = [this](Function &F) -> TargetLibraryInfo & { |
618 | return this->getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); |
619 | }; |
620 | |
621 | const auto *TM = &getAnalysis<TargetPassConfig>().getTM<TargetMachine>(); |
622 | PreISelIntrinsicLowering Lowering(TM, LookupTTI, LookupTLI); |
623 | return Lowering.lowerIntrinsics(M); |
624 | } |
625 | }; |
626 | |
627 | } // end anonymous namespace |
628 | |
629 | char PreISelIntrinsicLoweringLegacyPass::ID; |
630 | |
631 | INITIALIZE_PASS_BEGIN(PreISelIntrinsicLoweringLegacyPass, |
632 | "pre-isel-intrinsic-lowering" , |
633 | "Pre-ISel Intrinsic Lowering" , false, false) |
634 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) |
635 | INITIALIZE_PASS_DEPENDENCY(TargetPassConfig) |
636 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) |
637 | INITIALIZE_PASS_END(PreISelIntrinsicLoweringLegacyPass, |
638 | "pre-isel-intrinsic-lowering" , |
639 | "Pre-ISel Intrinsic Lowering" , false, false) |
640 | |
641 | ModulePass *llvm::createPreISelIntrinsicLoweringPass() { |
642 | return new PreISelIntrinsicLoweringLegacyPass(); |
643 | } |
644 | |
645 | PreservedAnalyses PreISelIntrinsicLoweringPass::run(Module &M, |
646 | ModuleAnalysisManager &AM) { |
647 | auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager(); |
648 | |
649 | auto LookupTTI = [&FAM](Function &F) -> TargetTransformInfo & { |
650 | return FAM.getResult<TargetIRAnalysis>(IR&: F); |
651 | }; |
652 | auto LookupTLI = [&FAM](Function &F) -> TargetLibraryInfo & { |
653 | return FAM.getResult<TargetLibraryAnalysis>(IR&: F); |
654 | }; |
655 | |
656 | PreISelIntrinsicLowering Lowering(TM, LookupTTI, LookupTLI); |
657 | if (!Lowering.lowerIntrinsics(M)) |
658 | return PreservedAnalyses::all(); |
659 | else |
660 | return PreservedAnalyses::none(); |
661 | } |
662 | |