1 #include "llvm/Transforms/Utils/VNCoercion.h"
2 #include "llvm/Analysis/ConstantFolding.h"
3 #include "llvm/Analysis/ValueTracking.h"
4 #include "llvm/IR/IRBuilder.h"
5 #include "llvm/IR/IntrinsicInst.h"
6 #include "llvm/Support/Debug.h"
7 
8 #define DEBUG_TYPE "vncoerce"
9 
10 namespace llvm {
11 namespace VNCoercion {
12 
13 static bool isFirstClassAggregateOrScalableType(Type *Ty) {
14   return Ty->isStructTy() || Ty->isArrayTy() || isa<ScalableVectorType>(Ty);
15 }
16 
17 /// Return true if coerceAvailableValueToLoadType will succeed.
18 bool canCoerceMustAliasedValueToLoad(Value *StoredVal, Type *LoadTy,
19                                      const DataLayout &DL) {
20   Type *StoredTy = StoredVal->getType();
21 
22   if (StoredTy == LoadTy)
23     return true;
24 
25   // If the loaded/stored value is a first class array/struct, or scalable type,
26   // don't try to transform them. We need to be able to bitcast to integer.
27   if (isFirstClassAggregateOrScalableType(LoadTy) ||
28       isFirstClassAggregateOrScalableType(StoredTy))
29     return false;
30 
31   uint64_t StoreSize = DL.getTypeSizeInBits(StoredTy).getFixedValue();
32 
33   // The store size must be byte-aligned to support future type casts.
34   if (llvm::alignTo(StoreSize, 8) != StoreSize)
35     return false;
36 
37   // The store has to be at least as big as the load.
38   if (StoreSize < DL.getTypeSizeInBits(LoadTy).getFixedValue())
39     return false;
40 
41   bool StoredNI = DL.isNonIntegralPointerType(StoredTy->getScalarType());
42   bool LoadNI = DL.isNonIntegralPointerType(LoadTy->getScalarType());
43   // Don't coerce non-integral pointers to integers or vice versa.
44   if (StoredNI != LoadNI) {
45     // As a special case, allow coercion of memset used to initialize
46     // an array w/null.  Despite non-integral pointers not generally having a
47     // specific bit pattern, we do assume null is zero.
48     if (auto *CI = dyn_cast<Constant>(StoredVal))
49       return CI->isNullValue();
50     return false;
51   } else if (StoredNI && LoadNI &&
52              StoredTy->getPointerAddressSpace() !=
53                  LoadTy->getPointerAddressSpace()) {
54     return false;
55   }
56 
57 
58   // The implementation below uses inttoptr for vectors of unequal size; we
59   // can't allow this for non integral pointers. We could teach it to extract
60   // exact subvectors if desired.
61   if (StoredNI && StoreSize != DL.getTypeSizeInBits(LoadTy).getFixedValue())
62     return false;
63 
64   if (StoredTy->isTargetExtTy() || LoadTy->isTargetExtTy())
65     return false;
66 
67   return true;
68 }
69 
70 /// If we saw a store of a value to memory, and
71 /// then a load from a must-aliased pointer of a different type, try to coerce
72 /// the stored value.  LoadedTy is the type of the load we want to replace.
73 /// IRB is IRBuilder used to insert new instructions.
74 ///
75 /// If we can't do it, return null.
76 Value *coerceAvailableValueToLoadType(Value *StoredVal, Type *LoadedTy,
77                                       IRBuilderBase &Helper,
78                                       const DataLayout &DL) {
79   assert(canCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL) &&
80          "precondition violation - materialization can't fail");
81   if (auto *C = dyn_cast<Constant>(StoredVal))
82     StoredVal = ConstantFoldConstant(C, DL);
83 
84   // If this is already the right type, just return it.
85   Type *StoredValTy = StoredVal->getType();
86 
87   uint64_t StoredValSize = DL.getTypeSizeInBits(StoredValTy).getFixedValue();
88   uint64_t LoadedValSize = DL.getTypeSizeInBits(LoadedTy).getFixedValue();
89 
90   // If the store and reload are the same size, we can always reuse it.
91   if (StoredValSize == LoadedValSize) {
92     // Pointer to Pointer -> use bitcast.
93     if (StoredValTy->isPtrOrPtrVectorTy() && LoadedTy->isPtrOrPtrVectorTy()) {
94       StoredVal = Helper.CreateBitCast(StoredVal, LoadedTy);
95     } else {
96       // Convert source pointers to integers, which can be bitcast.
97       if (StoredValTy->isPtrOrPtrVectorTy()) {
98         StoredValTy = DL.getIntPtrType(StoredValTy);
99         StoredVal = Helper.CreatePtrToInt(StoredVal, StoredValTy);
100       }
101 
102       Type *TypeToCastTo = LoadedTy;
103       if (TypeToCastTo->isPtrOrPtrVectorTy())
104         TypeToCastTo = DL.getIntPtrType(TypeToCastTo);
105 
106       if (StoredValTy != TypeToCastTo)
107         StoredVal = Helper.CreateBitCast(StoredVal, TypeToCastTo);
108 
109       // Cast to pointer if the load needs a pointer type.
110       if (LoadedTy->isPtrOrPtrVectorTy())
111         StoredVal = Helper.CreateIntToPtr(StoredVal, LoadedTy);
112     }
113 
114     if (auto *C = dyn_cast<ConstantExpr>(StoredVal))
115       StoredVal = ConstantFoldConstant(C, DL);
116 
117     return StoredVal;
118   }
119   // If the loaded value is smaller than the available value, then we can
120   // extract out a piece from it.  If the available value is too small, then we
121   // can't do anything.
122   assert(StoredValSize >= LoadedValSize &&
123          "canCoerceMustAliasedValueToLoad fail");
124 
125   // Convert source pointers to integers, which can be manipulated.
126   if (StoredValTy->isPtrOrPtrVectorTy()) {
127     StoredValTy = DL.getIntPtrType(StoredValTy);
128     StoredVal = Helper.CreatePtrToInt(StoredVal, StoredValTy);
129   }
130 
131   // Convert vectors and fp to integer, which can be manipulated.
132   if (!StoredValTy->isIntegerTy()) {
133     StoredValTy = IntegerType::get(StoredValTy->getContext(), StoredValSize);
134     StoredVal = Helper.CreateBitCast(StoredVal, StoredValTy);
135   }
136 
137   // If this is a big-endian system, we need to shift the value down to the low
138   // bits so that a truncate will work.
139   if (DL.isBigEndian()) {
140     uint64_t ShiftAmt = DL.getTypeStoreSizeInBits(StoredValTy).getFixedValue() -
141                         DL.getTypeStoreSizeInBits(LoadedTy).getFixedValue();
142     StoredVal = Helper.CreateLShr(
143         StoredVal, ConstantInt::get(StoredVal->getType(), ShiftAmt));
144   }
145 
146   // Truncate the integer to the right size now.
147   Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadedValSize);
148   StoredVal = Helper.CreateTruncOrBitCast(StoredVal, NewIntTy);
149 
150   if (LoadedTy != NewIntTy) {
151     // If the result is a pointer, inttoptr.
152     if (LoadedTy->isPtrOrPtrVectorTy())
153       StoredVal = Helper.CreateIntToPtr(StoredVal, LoadedTy);
154     else
155       // Otherwise, bitcast.
156       StoredVal = Helper.CreateBitCast(StoredVal, LoadedTy);
157   }
158 
159   if (auto *C = dyn_cast<Constant>(StoredVal))
160     StoredVal = ConstantFoldConstant(C, DL);
161 
162   return StoredVal;
163 }
164 
165 /// This function is called when we have a memdep query of a load that ends up
166 /// being a clobbering memory write (store, memset, memcpy, memmove).  This
167 /// means that the write *may* provide bits used by the load but we can't be
168 /// sure because the pointers don't must-alias.
169 ///
170 /// Check this case to see if there is anything more we can do before we give
171 /// up.  This returns -1 if we have to give up, or a byte number in the stored
172 /// value of the piece that feeds the load.
173 static int analyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
174                                           Value *WritePtr,
175                                           uint64_t WriteSizeInBits,
176                                           const DataLayout &DL) {
177   // If the loaded/stored value is a first class array/struct, or scalable type,
178   // don't try to transform them. We need to be able to bitcast to integer.
179   if (isFirstClassAggregateOrScalableType(LoadTy))
180     return -1;
181 
182   int64_t StoreOffset = 0, LoadOffset = 0;
183   Value *StoreBase =
184       GetPointerBaseWithConstantOffset(WritePtr, StoreOffset, DL);
185   Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, DL);
186   if (StoreBase != LoadBase)
187     return -1;
188 
189   uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy).getFixedValue();
190 
191   if ((WriteSizeInBits & 7) | (LoadSize & 7))
192     return -1;
193   uint64_t StoreSize = WriteSizeInBits / 8; // Convert to bytes.
194   LoadSize /= 8;
195 
196   // If the Load isn't completely contained within the stored bits, we don't
197   // have all the bits to feed it.  We could do something crazy in the future
198   // (issue a smaller load then merge the bits in) but this seems unlikely to be
199   // valuable.
200   if (StoreOffset > LoadOffset ||
201       StoreOffset + int64_t(StoreSize) < LoadOffset + int64_t(LoadSize))
202     return -1;
203 
204   // Okay, we can do this transformation.  Return the number of bytes into the
205   // store that the load is.
206   return LoadOffset - StoreOffset;
207 }
208 
209 /// This function is called when we have a
210 /// memdep query of a load that ends up being a clobbering store.
211 int analyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
212                                    StoreInst *DepSI, const DataLayout &DL) {
213   auto *StoredVal = DepSI->getValueOperand();
214 
215   // Cannot handle reading from store of first-class aggregate or scalable type.
216   if (isFirstClassAggregateOrScalableType(StoredVal->getType()))
217     return -1;
218 
219   if (!canCoerceMustAliasedValueToLoad(StoredVal, LoadTy, DL))
220     return -1;
221 
222   Value *StorePtr = DepSI->getPointerOperand();
223   uint64_t StoreSize =
224       DL.getTypeSizeInBits(DepSI->getValueOperand()->getType()).getFixedValue();
225   return analyzeLoadFromClobberingWrite(LoadTy, LoadPtr, StorePtr, StoreSize,
226                                         DL);
227 }
228 
229 /// This function is called when we have a
230 /// memdep query of a load that ends up being clobbered by another load.  See if
231 /// the other load can feed into the second load.
232 int analyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, LoadInst *DepLI,
233                                   const DataLayout &DL) {
234   // Cannot handle reading from store of first-class aggregate yet.
235   if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
236     return -1;
237 
238   if (!canCoerceMustAliasedValueToLoad(DepLI, LoadTy, DL))
239     return -1;
240 
241   Value *DepPtr = DepLI->getPointerOperand();
242   uint64_t DepSize = DL.getTypeSizeInBits(DepLI->getType()).getFixedValue();
243   return analyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, DL);
244 }
245 
246 int analyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
247                                      MemIntrinsic *MI, const DataLayout &DL) {
248   // If the mem operation is a non-constant size, we can't handle it.
249   ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
250   if (!SizeCst)
251     return -1;
252   uint64_t MemSizeInBits = SizeCst->getZExtValue() * 8;
253 
254   // If this is memset, we just need to see if the offset is valid in the size
255   // of the memset..
256   if (const auto *memset_inst = dyn_cast<MemSetInst>(MI)) {
257     if (DL.isNonIntegralPointerType(LoadTy->getScalarType())) {
258       auto *CI = dyn_cast<ConstantInt>(memset_inst->getValue());
259       if (!CI || !CI->isZero())
260         return -1;
261     }
262     return analyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
263                                           MemSizeInBits, DL);
264   }
265 
266   // If we have a memcpy/memmove, the only case we can handle is if this is a
267   // copy from constant memory.  In that case, we can read directly from the
268   // constant memory.
269   MemTransferInst *MTI = cast<MemTransferInst>(MI);
270 
271   Constant *Src = dyn_cast<Constant>(MTI->getSource());
272   if (!Src)
273     return -1;
274 
275   GlobalVariable *GV = dyn_cast<GlobalVariable>(getUnderlyingObject(Src));
276   if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
277     return -1;
278 
279   // See if the access is within the bounds of the transfer.
280   int Offset = analyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
281                                               MemSizeInBits, DL);
282   if (Offset == -1)
283     return Offset;
284 
285   // Otherwise, see if we can constant fold a load from the constant with the
286   // offset applied as appropriate.
287   unsigned IndexSize = DL.getIndexTypeSizeInBits(Src->getType());
288   if (ConstantFoldLoadFromConstPtr(Src, LoadTy, APInt(IndexSize, Offset), DL))
289     return Offset;
290   return -1;
291 }
292 
293 static Value *getStoreValueForLoadHelper(Value *SrcVal, unsigned Offset,
294                                          Type *LoadTy, IRBuilderBase &Builder,
295                                          const DataLayout &DL) {
296   LLVMContext &Ctx = SrcVal->getType()->getContext();
297 
298   // If two pointers are in the same address space, they have the same size,
299   // so we don't need to do any truncation, etc. This avoids introducing
300   // ptrtoint instructions for pointers that may be non-integral.
301   if (SrcVal->getType()->isPointerTy() && LoadTy->isPointerTy() &&
302       cast<PointerType>(SrcVal->getType())->getAddressSpace() ==
303           cast<PointerType>(LoadTy)->getAddressSpace()) {
304     return SrcVal;
305   }
306 
307   uint64_t StoreSize =
308       (DL.getTypeSizeInBits(SrcVal->getType()).getFixedValue() + 7) / 8;
309   uint64_t LoadSize = (DL.getTypeSizeInBits(LoadTy).getFixedValue() + 7) / 8;
310   // Compute which bits of the stored value are being used by the load.  Convert
311   // to an integer type to start with.
312   if (SrcVal->getType()->isPtrOrPtrVectorTy())
313     SrcVal =
314         Builder.CreatePtrToInt(SrcVal, DL.getIntPtrType(SrcVal->getType()));
315   if (!SrcVal->getType()->isIntegerTy())
316     SrcVal =
317         Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize * 8));
318 
319   // Shift the bits to the least significant depending on endianness.
320   unsigned ShiftAmt;
321   if (DL.isLittleEndian())
322     ShiftAmt = Offset * 8;
323   else
324     ShiftAmt = (StoreSize - LoadSize - Offset) * 8;
325   if (ShiftAmt)
326     SrcVal = Builder.CreateLShr(SrcVal,
327                                 ConstantInt::get(SrcVal->getType(), ShiftAmt));
328 
329   if (LoadSize != StoreSize)
330     SrcVal = Builder.CreateTruncOrBitCast(SrcVal,
331                                           IntegerType::get(Ctx, LoadSize * 8));
332   return SrcVal;
333 }
334 
335 Value *getValueForLoad(Value *SrcVal, unsigned Offset, Type *LoadTy,
336                        Instruction *InsertPt, const DataLayout &DL) {
337 
338 #ifndef NDEBUG
339   unsigned SrcValSize = DL.getTypeStoreSize(SrcVal->getType()).getFixedValue();
340   unsigned LoadSize = DL.getTypeStoreSize(LoadTy).getFixedValue();
341   assert(Offset + LoadSize <= SrcValSize);
342 #endif
343   IRBuilder<> Builder(InsertPt);
344   SrcVal = getStoreValueForLoadHelper(SrcVal, Offset, LoadTy, Builder, DL);
345   return coerceAvailableValueToLoadType(SrcVal, LoadTy, Builder, DL);
346 }
347 
348 Constant *getConstantValueForLoad(Constant *SrcVal, unsigned Offset,
349                                   Type *LoadTy, const DataLayout &DL) {
350 #ifndef NDEBUG
351   unsigned SrcValSize = DL.getTypeStoreSize(SrcVal->getType()).getFixedValue();
352   unsigned LoadSize = DL.getTypeStoreSize(LoadTy).getFixedValue();
353   assert(Offset + LoadSize <= SrcValSize);
354 #endif
355   return ConstantFoldLoadFromConst(SrcVal, LoadTy, APInt(32, Offset), DL);
356 }
357 
358 /// This function is called when we have a
359 /// memdep query of a load that ends up being a clobbering mem intrinsic.
360 Value *getMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
361                               Type *LoadTy, Instruction *InsertPt,
362                               const DataLayout &DL) {
363   LLVMContext &Ctx = LoadTy->getContext();
364   uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy).getFixedValue() / 8;
365   IRBuilder<> Builder(InsertPt);
366 
367   // We know that this method is only called when the mem transfer fully
368   // provides the bits for the load.
369   if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
370     // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
371     // independently of what the offset is.
372     Value *Val = MSI->getValue();
373     if (LoadSize != 1)
374       Val =
375           Builder.CreateZExtOrBitCast(Val, IntegerType::get(Ctx, LoadSize * 8));
376     Value *OneElt = Val;
377 
378     // Splat the value out to the right number of bits.
379     for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize;) {
380       // If we can double the number of bytes set, do it.
381       if (NumBytesSet * 2 <= LoadSize) {
382         Value *ShVal = Builder.CreateShl(
383             Val, ConstantInt::get(Val->getType(), NumBytesSet * 8));
384         Val = Builder.CreateOr(Val, ShVal);
385         NumBytesSet <<= 1;
386         continue;
387       }
388 
389       // Otherwise insert one byte at a time.
390       Value *ShVal =
391           Builder.CreateShl(Val, ConstantInt::get(Val->getType(), 1 * 8));
392       Val = Builder.CreateOr(OneElt, ShVal);
393       ++NumBytesSet;
394     }
395 
396     return coerceAvailableValueToLoadType(Val, LoadTy, Builder, DL);
397   }
398 
399   // Otherwise, this is a memcpy/memmove from a constant global.
400   MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
401   Constant *Src = cast<Constant>(MTI->getSource());
402   unsigned IndexSize = DL.getIndexTypeSizeInBits(Src->getType());
403   return ConstantFoldLoadFromConstPtr(Src, LoadTy, APInt(IndexSize, Offset),
404                                       DL);
405 }
406 
407 Constant *getConstantMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
408                                          Type *LoadTy, const DataLayout &DL) {
409   LLVMContext &Ctx = LoadTy->getContext();
410   uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy).getFixedValue() / 8;
411 
412   // We know that this method is only called when the mem transfer fully
413   // provides the bits for the load.
414   if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
415     auto *Val = dyn_cast<ConstantInt>(MSI->getValue());
416     if (!Val)
417       return nullptr;
418 
419     Val = ConstantInt::get(Ctx, APInt::getSplat(LoadSize * 8, Val->getValue()));
420     return ConstantFoldLoadFromConst(Val, LoadTy, DL);
421   }
422 
423   // Otherwise, this is a memcpy/memmove from a constant global.
424   MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
425   Constant *Src = cast<Constant>(MTI->getSource());
426   unsigned IndexSize = DL.getIndexTypeSizeInBits(Src->getType());
427   return ConstantFoldLoadFromConstPtr(Src, LoadTy, APInt(IndexSize, Offset),
428                                       DL);
429 }
430 } // namespace VNCoercion
431 } // namespace llvm
432