Move EVER MORE stuff over to LLVMContext.
[oota-llvm.git] / lib / CodeGen / IntrinsicLowering.cpp
1 //===-- IntrinsicLowering.cpp - Intrinsic Lowering default implementation -===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the IntrinsicLowering class.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "llvm/Constants.h"
15 #include "llvm/DerivedTypes.h"
16 #include "llvm/Module.h"
17 #include "llvm/Type.h"
18 #include "llvm/CodeGen/IntrinsicLowering.h"
19 #include "llvm/Support/IRBuilder.h"
20 #include "llvm/Support/ErrorHandling.h"
21 #include "llvm/Target/TargetData.h"
22 #include "llvm/ADT/SmallVector.h"
23 using namespace llvm;
24
25 template <class ArgIt>
26 static void EnsureFunctionExists(Module &M, const char *Name,
27                                  ArgIt ArgBegin, ArgIt ArgEnd,
28                                  const Type *RetTy) {
29   // Insert a correctly-typed definition now.
30   std::vector<const Type *> ParamTys;
31   for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
32     ParamTys.push_back(I->getType());
33   M.getOrInsertFunction(Name, FunctionType::get(RetTy, ParamTys, false));
34 }
35
36 static void EnsureFPIntrinsicsExist(Module &M, Function *Fn,
37                                     const char *FName,
38                                     const char *DName, const char *LDName) {
39   // Insert definitions for all the floating point types.
40   switch((int)Fn->arg_begin()->getType()->getTypeID()) {
41   case Type::FloatTyID:
42     EnsureFunctionExists(M, FName, Fn->arg_begin(), Fn->arg_end(),
43                          Type::FloatTy);
44     break;
45   case Type::DoubleTyID:
46     EnsureFunctionExists(M, DName, Fn->arg_begin(), Fn->arg_end(),
47                          Type::DoubleTy);
48     break;
49   case Type::X86_FP80TyID:
50   case Type::FP128TyID:
51   case Type::PPC_FP128TyID:
52     EnsureFunctionExists(M, LDName, Fn->arg_begin(), Fn->arg_end(),
53                          Fn->arg_begin()->getType());
54     break;
55   }
56 }
57
58 /// ReplaceCallWith - This function is used when we want to lower an intrinsic
59 /// call to a call of an external function.  This handles hard cases such as
60 /// when there was already a prototype for the external function, and if that
61 /// prototype doesn't match the arguments we expect to pass in.
62 template <class ArgIt>
63 static CallInst *ReplaceCallWith(const char *NewFn, CallInst *CI,
64                                  ArgIt ArgBegin, ArgIt ArgEnd,
65                                  const Type *RetTy) {
66   // If we haven't already looked up this function, check to see if the
67   // program already contains a function with this name.
68   Module *M = CI->getParent()->getParent()->getParent();
69   // Get or insert the definition now.
70   std::vector<const Type *> ParamTys;
71   for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
72     ParamTys.push_back((*I)->getType());
73   Constant* FCache = M->getOrInsertFunction(NewFn,
74                                   FunctionType::get(RetTy, ParamTys, false));
75
76   IRBuilder<> Builder(CI->getParent(), CI);
77   SmallVector<Value *, 8> Args(ArgBegin, ArgEnd);
78   CallInst *NewCI = Builder.CreateCall(FCache, Args.begin(), Args.end());
79   NewCI->setName(CI->getName());
80   if (!CI->use_empty())
81     CI->replaceAllUsesWith(NewCI);
82   return NewCI;
83 }
84
85 void IntrinsicLowering::AddPrototypes(Module &M) {
86   for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
87     if (I->isDeclaration() && !I->use_empty())
88       switch (I->getIntrinsicID()) {
89       default: break;
90       case Intrinsic::setjmp:
91         EnsureFunctionExists(M, "setjmp", I->arg_begin(), I->arg_end(),
92                              Type::Int32Ty);
93         break;
94       case Intrinsic::longjmp:
95         EnsureFunctionExists(M, "longjmp", I->arg_begin(), I->arg_end(),
96                              Type::VoidTy);
97         break;
98       case Intrinsic::siglongjmp:
99         EnsureFunctionExists(M, "abort", I->arg_end(), I->arg_end(),
100                              Type::VoidTy);
101         break;
102       case Intrinsic::memcpy:
103         M.getOrInsertFunction("memcpy", PointerType::getUnqual(Type::Int8Ty),
104                               PointerType::getUnqual(Type::Int8Ty), 
105                               PointerType::getUnqual(Type::Int8Ty), 
106                               TD.getIntPtrType(), (Type *)0);
107         break;
108       case Intrinsic::memmove:
109         M.getOrInsertFunction("memmove", PointerType::getUnqual(Type::Int8Ty),
110                               PointerType::getUnqual(Type::Int8Ty), 
111                               PointerType::getUnqual(Type::Int8Ty), 
112                               TD.getIntPtrType(), (Type *)0);
113         break;
114       case Intrinsic::memset:
115         M.getOrInsertFunction("memset", PointerType::getUnqual(Type::Int8Ty),
116                               PointerType::getUnqual(Type::Int8Ty), 
117                               Type::Int32Ty, 
118                               TD.getIntPtrType(), (Type *)0);
119         break;
120       case Intrinsic::sqrt:
121         EnsureFPIntrinsicsExist(M, I, "sqrtf", "sqrt", "sqrtl");
122         break;
123       case Intrinsic::sin:
124         EnsureFPIntrinsicsExist(M, I, "sinf", "sin", "sinl");
125         break;
126       case Intrinsic::cos:
127         EnsureFPIntrinsicsExist(M, I, "cosf", "cos", "cosl");
128         break;
129       case Intrinsic::pow:
130         EnsureFPIntrinsicsExist(M, I, "powf", "pow", "powl");
131         break;
132       case Intrinsic::log:
133         EnsureFPIntrinsicsExist(M, I, "logf", "log", "logl");
134         break;
135       case Intrinsic::log2:
136         EnsureFPIntrinsicsExist(M, I, "log2f", "log2", "log2l");
137         break;
138       case Intrinsic::log10:
139         EnsureFPIntrinsicsExist(M, I, "log10f", "log10", "log10l");
140         break;
141       case Intrinsic::exp:
142         EnsureFPIntrinsicsExist(M, I, "expf", "exp", "expl");
143         break;
144       case Intrinsic::exp2:
145         EnsureFPIntrinsicsExist(M, I, "exp2f", "exp2", "exp2l");
146         break;
147       }
148 }
149
150 /// LowerBSWAP - Emit the code to lower bswap of V before the specified
151 /// instruction IP.
152 static Value *LowerBSWAP(LLVMContext &Context, Value *V, Instruction *IP) {
153   assert(V->getType()->isInteger() && "Can't bswap a non-integer type!");
154
155   unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
156   
157   IRBuilder<> Builder(IP->getParent(), IP);
158
159   switch(BitSize) {
160   default: llvm_unreachable("Unhandled type size of value to byteswap!");
161   case 16: {
162     Value *Tmp1 = Builder.CreateShl(V, Context.getConstantInt(V->getType(), 8),
163                                     "bswap.2");
164     Value *Tmp2 = Builder.CreateLShr(V, Context.getConstantInt(V->getType(), 8),
165                                      "bswap.1");
166     V = Builder.CreateOr(Tmp1, Tmp2, "bswap.i16");
167     break;
168   }
169   case 32: {
170     Value *Tmp4 = Builder.CreateShl(V, Context.getConstantInt(V->getType(), 24),
171                                     "bswap.4");
172     Value *Tmp3 = Builder.CreateShl(V, Context.getConstantInt(V->getType(), 8),
173                                     "bswap.3");
174     Value *Tmp2 = Builder.CreateLShr(V, Context.getConstantInt(V->getType(), 8),
175                                      "bswap.2");
176     Value *Tmp1 = Builder.CreateLShr(V,Context.getConstantInt(V->getType(), 24),
177                                      "bswap.1");
178     Tmp3 = Builder.CreateAnd(Tmp3,
179                              Context.getConstantInt(Type::Int32Ty, 0xFF0000),
180                              "bswap.and3");
181     Tmp2 = Builder.CreateAnd(Tmp2,
182                              Context.getConstantInt(Type::Int32Ty, 0xFF00),
183                              "bswap.and2");
184     Tmp4 = Builder.CreateOr(Tmp4, Tmp3, "bswap.or1");
185     Tmp2 = Builder.CreateOr(Tmp2, Tmp1, "bswap.or2");
186     V = Builder.CreateOr(Tmp4, Tmp2, "bswap.i32");
187     break;
188   }
189   case 64: {
190     Value *Tmp8 = Builder.CreateShl(V, Context.getConstantInt(V->getType(), 56),
191                                     "bswap.8");
192     Value *Tmp7 = Builder.CreateShl(V, Context.getConstantInt(V->getType(), 40),
193                                     "bswap.7");
194     Value *Tmp6 = Builder.CreateShl(V, Context.getConstantInt(V->getType(), 24),
195                                     "bswap.6");
196     Value *Tmp5 = Builder.CreateShl(V, Context.getConstantInt(V->getType(), 8),
197                                     "bswap.5");
198     Value* Tmp4 = Builder.CreateLShr(V, Context.getConstantInt(V->getType(), 8),
199                                      "bswap.4");
200     Value* Tmp3 = Builder.CreateLShr(V, 
201                                      Context.getConstantInt(V->getType(), 24),
202                                      "bswap.3");
203     Value* Tmp2 = Builder.CreateLShr(V, 
204                                      Context.getConstantInt(V->getType(), 40),
205                                      "bswap.2");
206     Value* Tmp1 = Builder.CreateLShr(V, 
207                                      Context.getConstantInt(V->getType(), 56),
208                                      "bswap.1");
209     Tmp7 = Builder.CreateAnd(Tmp7,
210                              Context.getConstantInt(Type::Int64Ty,
211                                               0xFF000000000000ULL),
212                              "bswap.and7");
213     Tmp6 = Builder.CreateAnd(Tmp6,
214                              Context.getConstantInt(Type::Int64Ty,
215                                               0xFF0000000000ULL),
216                              "bswap.and6");
217     Tmp5 = Builder.CreateAnd(Tmp5,
218                         Context.getConstantInt(Type::Int64Ty, 0xFF00000000ULL),
219                              "bswap.and5");
220     Tmp4 = Builder.CreateAnd(Tmp4,
221                         Context.getConstantInt(Type::Int64Ty, 0xFF000000ULL),
222                              "bswap.and4");
223     Tmp3 = Builder.CreateAnd(Tmp3,
224                              Context.getConstantInt(Type::Int64Ty, 0xFF0000ULL),
225                              "bswap.and3");
226     Tmp2 = Builder.CreateAnd(Tmp2,
227                              Context.getConstantInt(Type::Int64Ty, 0xFF00ULL),
228                              "bswap.and2");
229     Tmp8 = Builder.CreateOr(Tmp8, Tmp7, "bswap.or1");
230     Tmp6 = Builder.CreateOr(Tmp6, Tmp5, "bswap.or2");
231     Tmp4 = Builder.CreateOr(Tmp4, Tmp3, "bswap.or3");
232     Tmp2 = Builder.CreateOr(Tmp2, Tmp1, "bswap.or4");
233     Tmp8 = Builder.CreateOr(Tmp8, Tmp6, "bswap.or5");
234     Tmp4 = Builder.CreateOr(Tmp4, Tmp2, "bswap.or6");
235     V = Builder.CreateOr(Tmp8, Tmp4, "bswap.i64");
236     break;
237   }
238   }
239   return V;
240 }
241
242 /// LowerCTPOP - Emit the code to lower ctpop of V before the specified
243 /// instruction IP.
244 static Value *LowerCTPOP(LLVMContext &Context, Value *V, Instruction *IP) {
245   assert(V->getType()->isInteger() && "Can't ctpop a non-integer type!");
246
247   static const uint64_t MaskValues[6] = {
248     0x5555555555555555ULL, 0x3333333333333333ULL,
249     0x0F0F0F0F0F0F0F0FULL, 0x00FF00FF00FF00FFULL,
250     0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL
251   };
252
253   IRBuilder<> Builder(IP->getParent(), IP);
254
255   unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
256   unsigned WordSize = (BitSize + 63) / 64;
257   Value *Count = Context.getConstantInt(V->getType(), 0);
258
259   for (unsigned n = 0; n < WordSize; ++n) {
260     Value *PartValue = V;
261     for (unsigned i = 1, ct = 0; i < (BitSize>64 ? 64 : BitSize); 
262          i <<= 1, ++ct) {
263       Value *MaskCst = Context.getConstantInt(V->getType(), MaskValues[ct]);
264       Value *LHS = Builder.CreateAnd(PartValue, MaskCst, "cppop.and1");
265       Value *VShift = Builder.CreateLShr(PartValue,
266                                         Context.getConstantInt(V->getType(), i),
267                                          "ctpop.sh");
268       Value *RHS = Builder.CreateAnd(VShift, MaskCst, "cppop.and2");
269       PartValue = Builder.CreateAdd(LHS, RHS, "ctpop.step");
270     }
271     Count = Builder.CreateAdd(PartValue, Count, "ctpop.part");
272     if (BitSize > 64) {
273       V = Builder.CreateLShr(V, Context.getConstantInt(V->getType(), 64),
274                              "ctpop.part.sh");
275       BitSize -= 64;
276     }
277   }
278
279   return Count;
280 }
281
282 /// LowerCTLZ - Emit the code to lower ctlz of V before the specified
283 /// instruction IP.
284 static Value *LowerCTLZ(LLVMContext &Context, Value *V, Instruction *IP) {
285
286   IRBuilder<> Builder(IP->getParent(), IP);
287
288   unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
289   for (unsigned i = 1; i < BitSize; i <<= 1) {
290     Value *ShVal = Context.getConstantInt(V->getType(), i);
291     ShVal = Builder.CreateLShr(V, ShVal, "ctlz.sh");
292     V = Builder.CreateOr(V, ShVal, "ctlz.step");
293   }
294
295   V = Builder.CreateNot(V);
296   return LowerCTPOP(Context, V, IP);
297 }
298
299 static void ReplaceFPIntrinsicWithCall(CallInst *CI, const char *Fname,
300                                        const char *Dname,
301                                        const char *LDname) {
302   switch (CI->getOperand(1)->getType()->getTypeID()) {
303   default: llvm_unreachable("Invalid type in intrinsic");
304   case Type::FloatTyID:
305     ReplaceCallWith(Fname, CI, CI->op_begin() + 1, CI->op_end(),
306                   Type::FloatTy);
307     break;
308   case Type::DoubleTyID:
309     ReplaceCallWith(Dname, CI, CI->op_begin() + 1, CI->op_end(),
310                   Type::DoubleTy);
311     break;
312   case Type::X86_FP80TyID:
313   case Type::FP128TyID:
314   case Type::PPC_FP128TyID:
315     ReplaceCallWith(LDname, CI, CI->op_begin() + 1, CI->op_end(),
316                   CI->getOperand(1)->getType());
317     break;
318   }
319 }
320
321 void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
322   IRBuilder<> Builder(CI->getParent(), CI);
323   LLVMContext *Context = CI->getParent()->getContext();
324
325   Function *Callee = CI->getCalledFunction();
326   assert(Callee && "Cannot lower an indirect call!");
327
328   switch (Callee->getIntrinsicID()) {
329   case Intrinsic::not_intrinsic:
330     llvm_report_error("Cannot lower a call to a non-intrinsic function '"+
331                       Callee->getName() + "'!");
332   default:
333     llvm_report_error("Code generator does not support intrinsic function '"+
334                       Callee->getName()+"'!");
335
336     // The setjmp/longjmp intrinsics should only exist in the code if it was
337     // never optimized (ie, right out of the CFE), or if it has been hacked on
338     // by the lowerinvoke pass.  In both cases, the right thing to do is to
339     // convert the call to an explicit setjmp or longjmp call.
340   case Intrinsic::setjmp: {
341     Value *V = ReplaceCallWith("setjmp", CI, CI->op_begin() + 1, CI->op_end(),
342                                Type::Int32Ty);
343     if (CI->getType() != Type::VoidTy)
344       CI->replaceAllUsesWith(V);
345     break;
346   }
347   case Intrinsic::sigsetjmp:
348      if (CI->getType() != Type::VoidTy)
349        CI->replaceAllUsesWith(Context->getNullValue(CI->getType()));
350      break;
351
352   case Intrinsic::longjmp: {
353     ReplaceCallWith("longjmp", CI, CI->op_begin() + 1, CI->op_end(),
354                     Type::VoidTy);
355     break;
356   }
357
358   case Intrinsic::siglongjmp: {
359     // Insert the call to abort
360     ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(), 
361                     Type::VoidTy);
362     break;
363   }
364   case Intrinsic::ctpop:
365     CI->replaceAllUsesWith(LowerCTPOP(*Context, CI->getOperand(1), CI));
366     break;
367
368   case Intrinsic::bswap:
369     CI->replaceAllUsesWith(LowerBSWAP(*Context, CI->getOperand(1), CI));
370     break;
371     
372   case Intrinsic::ctlz:
373     CI->replaceAllUsesWith(LowerCTLZ(*Context, CI->getOperand(1), CI));
374     break;
375
376   case Intrinsic::cttz: {
377     // cttz(x) -> ctpop(~X & (X-1))
378     Value *Src = CI->getOperand(1);
379     Value *NotSrc = Builder.CreateNot(Src);
380     NotSrc->setName(Src->getName() + ".not");
381     Value *SrcM1 = Context->getConstantInt(Src->getType(), 1);
382     SrcM1 = Builder.CreateSub(Src, SrcM1);
383     Src = LowerCTPOP(*Context, Builder.CreateAnd(NotSrc, SrcM1), CI);
384     CI->replaceAllUsesWith(Src);
385     break;
386   }
387
388   case Intrinsic::stacksave:
389   case Intrinsic::stackrestore: {
390     if (!Warned)
391       cerr << "WARNING: this target does not support the llvm.stack"
392            << (Callee->getIntrinsicID() == Intrinsic::stacksave ?
393                "save" : "restore") << " intrinsic.\n";
394     Warned = true;
395     if (Callee->getIntrinsicID() == Intrinsic::stacksave)
396       CI->replaceAllUsesWith(Context->getNullValue(CI->getType()));
397     break;
398   }
399     
400   case Intrinsic::returnaddress:
401   case Intrinsic::frameaddress:
402     cerr << "WARNING: this target does not support the llvm."
403          << (Callee->getIntrinsicID() == Intrinsic::returnaddress ?
404              "return" : "frame") << "address intrinsic.\n";
405     CI->replaceAllUsesWith(ConstantPointerNull::get(
406                                             cast<PointerType>(CI->getType())));
407     break;
408
409   case Intrinsic::prefetch:
410     break;    // Simply strip out prefetches on unsupported architectures
411
412   case Intrinsic::pcmarker:
413     break;    // Simply strip out pcmarker on unsupported architectures
414   case Intrinsic::readcyclecounter: {
415     cerr << "WARNING: this target does not support the llvm.readcyclecoun"
416          << "ter intrinsic.  It is being lowered to a constant 0\n";
417     CI->replaceAllUsesWith(Context->getConstantInt(Type::Int64Ty, 0));
418     break;
419   }
420
421   case Intrinsic::dbg_stoppoint:
422   case Intrinsic::dbg_region_start:
423   case Intrinsic::dbg_region_end:
424   case Intrinsic::dbg_func_start:
425   case Intrinsic::dbg_declare:
426     break;    // Simply strip out debugging intrinsics
427
428   case Intrinsic::eh_exception:
429   case Intrinsic::eh_selector_i32:
430   case Intrinsic::eh_selector_i64:
431     CI->replaceAllUsesWith(Context->getNullValue(CI->getType()));
432     break;
433
434   case Intrinsic::eh_typeid_for_i32:
435   case Intrinsic::eh_typeid_for_i64:
436     // Return something different to eh_selector.
437     CI->replaceAllUsesWith(Context->getConstantInt(CI->getType(), 1));
438     break;
439
440   case Intrinsic::var_annotation:
441     break;   // Strip out annotate intrinsic
442     
443   case Intrinsic::memcpy: {
444     const IntegerType *IntPtr = TD.getIntPtrType();
445     Value *Size = Builder.CreateIntCast(CI->getOperand(3), IntPtr,
446                                         /* isSigned */ false);
447     Value *Ops[3];
448     Ops[0] = CI->getOperand(1);
449     Ops[1] = CI->getOperand(2);
450     Ops[2] = Size;
451     ReplaceCallWith("memcpy", CI, Ops, Ops+3, CI->getOperand(1)->getType());
452     break;
453   }
454   case Intrinsic::memmove: {
455     const IntegerType *IntPtr = TD.getIntPtrType();
456     Value *Size = Builder.CreateIntCast(CI->getOperand(3), IntPtr,
457                                         /* isSigned */ false);
458     Value *Ops[3];
459     Ops[0] = CI->getOperand(1);
460     Ops[1] = CI->getOperand(2);
461     Ops[2] = Size;
462     ReplaceCallWith("memmove", CI, Ops, Ops+3, CI->getOperand(1)->getType());
463     break;
464   }
465   case Intrinsic::memset: {
466     const IntegerType *IntPtr = TD.getIntPtrType();
467     Value *Size = Builder.CreateIntCast(CI->getOperand(3), IntPtr,
468                                         /* isSigned */ false);
469     Value *Ops[3];
470     Ops[0] = CI->getOperand(1);
471     // Extend the amount to i32.
472     Ops[1] = Builder.CreateIntCast(CI->getOperand(2), Type::Int32Ty,
473                                    /* isSigned */ false);
474     Ops[2] = Size;
475     ReplaceCallWith("memset", CI, Ops, Ops+3, CI->getOperand(1)->getType());
476     break;
477   }
478   case Intrinsic::sqrt: {
479     ReplaceFPIntrinsicWithCall(CI, "sqrtf", "sqrt", "sqrtl");
480     break;
481   }
482   case Intrinsic::log: {
483     ReplaceFPIntrinsicWithCall(CI, "logf", "log", "logl");
484     break;
485   }
486   case Intrinsic::log2: {
487     ReplaceFPIntrinsicWithCall(CI, "log2f", "log2", "log2l");
488     break;
489   }
490   case Intrinsic::log10: {
491     ReplaceFPIntrinsicWithCall(CI, "log10f", "log10", "log10l");
492     break;
493   }
494   case Intrinsic::exp: {
495     ReplaceFPIntrinsicWithCall(CI, "expf", "exp", "expl");
496     break;
497   }
498   case Intrinsic::exp2: {
499     ReplaceFPIntrinsicWithCall(CI, "exp2f", "exp2", "exp2l");
500     break;
501   }
502   case Intrinsic::pow: {
503     ReplaceFPIntrinsicWithCall(CI, "powf", "pow", "powl");
504     break;
505   }
506   case Intrinsic::flt_rounds:
507      // Lower to "round to the nearest"
508      if (CI->getType() != Type::VoidTy)
509        CI->replaceAllUsesWith(Context->getConstantInt(CI->getType(), 1));
510      break;
511   }
512
513   assert(CI->use_empty() &&
514          "Lowering should have eliminated any uses of the intrinsic call!");
515   CI->eraseFromParent();
516 }