Revert of 27b6b89.

Change-Id: Ifb907ae61f2cb1eb151ba3cd9de7d2e0d2a577a4

Author: krystian-andrzejewski

IGC/Compiler/CISACodeGen/VectorPreProcess.cpp

@@ -41,7 +41,7 @@ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 using namespace llvm;
 using namespace IGC;
 
-//
+// 
 // Description of VectorPreProcess Pass
 //   The purpose is both to legalize vector types and to reduce register
 //   presure. Once this pass is done, there is no 3-element vector whose
@@ -53,7 +53,7 @@ using namespace IGC;
 //    are either multiple of DW, vector3, or their size is less than
 //    4 bytes (see details in code).  Vector3 will be specially
 //    handled later.
-//    For example,
+//    For example,  
 //        <16xi64> ---> four <4xi64>
 //        <15xi32> ---> <8xi32>, <7xi32>
 //        <13xi32> ---> <8xi32>, <5xi32>
@@ -62,7 +62,7 @@ using namespace IGC;
 //        <39xi8>  ---> <32xi8>, <4xi8>, <3xi8>
 //    Note that splitting keeps the vector element's type without
 //    changing it.
-//
+//        
 // 2. Special processing of 3-element vectors
 //    If (vector element's size < 4 bytes)
 //    {
@@ -91,11 +91,11 @@ using namespace IGC;
 //
 namespace
 {
-    // AbstractLoadInst and AbstractStoreInst abstract away the differences
+    // AbstractLoadInst and AbstractStoreInst abstract away the differences 
     // between ldraw and Load and between storeraw and Store.
     // Note on usage: The Value* passed as the ptr paramter to the Create method
-    // should be either the result of the getPointerOperand() method or the
-    // CreateConstScalarGEP() method. Do not attempt to do arithmetic
+    // should be either the result of the getPointerOperand() method or the 
+    // CreateConstScalarGEP() method. Do not attempt to do arithmetic 
     // (or pointer arithmetic) on these values.
     class AbstractLoadInst
     {
@@ -223,7 +223,7 @@ namespace
         }
         unsigned int getAlignment() const
         {
-            return isa<StoreInst>(m_inst) ? getStore()->getAlignment() : (unsigned int)llvm::cast<ConstantInt>(getStoreRaw()->getArgOperand(3))->getZExtValue();
+            return isa<StoreInst>(m_inst) ? getStore()->getAlignment() : getStoreRaw()->getArgOperand(2)->getType()->getPrimitiveSizeInBits() / 8;
         }
         void setAlignment(unsigned int alignment)
         {
@@ -430,8 +430,8 @@ bool VectorPreProcess::isValueUsedOnlyByEEI(Value *V, ExtractElementInst **EEIns
          UI != UE; ++UI )
     {
         ExtractElementInst *EEI = dyn_cast<ExtractElementInst>(*UI);
-        if( !EEI ||
-            (EEI->getOperand(0) != V) ||
+        if( !EEI || 
+            (EEI->getOperand(0) != V) || 
             !isa<ConstantInt>(EEI->getOperand(1)) )
         {
             return false;
@@ -451,7 +451,7 @@ bool VectorPreProcess::isValueUsedOnlyByEEI(Value *V, ExtractElementInst **EEIns
 
 // SVals[0:NumElements] has all scalar elements of vector VI. This function
 // tries to replace all uses of VI with SVals[...] if possible, If not
-// possible, re-generate the vector from SVals at the BB of VI.
+// possible, re-generate the vector from SVals at the BB of VI.  
 //
 // This function also erase VI.
 void VectorPreProcess::replaceAllVectorUsesWithScalars(Instruction *VI, ValVector& SVals)
@@ -557,7 +557,7 @@ void VectorPreProcess::createSplitVectorTypes(
         ++j;
     }
 
-    // Sub-vectors are
+    // Sub-vectors are 
     //   1. ebytes >=4, the remaing is a single sub-vector; or
     //   2. ebytes < 4, the remaining is splitted into
     //        one sub-vector of multiple 4xebytes, and
@@ -694,7 +694,7 @@ bool VectorPreProcess::splitStore(AbstractStoreInst& ASI, V2SMap& vecToSubVec)
     bool IsVolatile = ASI.getIsVolatile();
     uint32_t eOffset = 0;
     uint32_t EBytes = int_cast<unsigned int>(m_DL->getTypeAllocSize(ETy));
-
+    
     for (uint32_t i = 0, subIdx = 0; i < len; ++i)
     {
         VectorType *VTy1 = dyn_cast<VectorType>(tys[i]);
@@ -947,7 +947,7 @@ bool VectorPreProcess::splitVector3LoadStore(Instruction *Inst)
                 Elt2 = ALI->Create(eTy, offsetAddr, newAlign, ALI->getIsVolatile());
             }
 
-            // A little optimization here
+            // A little optimization here 
             ExtractElementInst *EEInsts[3];
             for (int i = 0; i < 3; ++i)
             {
@@ -982,7 +982,7 @@ bool VectorPreProcess::splitVector3LoadStore(Instruction *Inst)
             ALI->getInst()->eraseFromParent();
         }
         else
-        {
+        {   
             Value *Ptr = ASI->getPointerOperand();
             // Split 3-element into 2-element + 1 scalar
             Type *newVTy = VectorType::get(eTy, 2);
@@ -1088,7 +1088,7 @@ void VectorPreProcess::getOrGenScalarValues(
         for (uint32_t i = 0; i < nelts; ++i)
         {
             scalars[i] = udv;
-        }
+        }       
     }
     else if (ConstantVector* CV = dyn_cast<ConstantVector>(VecVal))
     {
@@ -1148,7 +1148,7 @@ void VectorPreProcess::getOrGenScalarValues(
             // VecVal is an argument or constant
             inst_b = F.begin()->getFirstInsertionPt();
         }
-
+        
         IRBuilder<> Builder(&(*inst_b));
         for (uint32_t i = 0; i < nelts; ++i)
         {

IGC/Compiler/CISACodeGen/helper.cpp

@@ -304,8 +304,7 @@ Value* CreateStoreRawIntrinsic(StoreInst *inst, Instruction* bufPtr, Value* offs
     {
         bufPtr,
         offsetVal,
-        storeVal,
-        builder.getInt32(storeVal->getType()->getScalarSizeInBits() / 8)
+        storeVal
     };
     Value* st = builder.CreateCall(func, attr);
     return st;

IGC/Compiler/FixResourcePtr.cpp

@@ -357,8 +357,7 @@ Value* FixResourcePtr::CreateStoreIntrinsic(StoreInst *inst, Instruction* bufPtr
     {
         bufPtr,
         offsetVal,
-        storeVal,
-        builder->getInt32(storeVal->getType()->getScalarSizeInBits() / 8)
+        storeVal
     };
     Value* st = builder->CreateCall(l, attr);
     return st;

IGC/Compiler/PromoteResourceToDirectAS.cpp

@@ -117,7 +117,7 @@ Value* GetBufferOperand(Instruction* inst)
             case GenISAIntrinsic::GenISA_floatatomicraw:
             case GenISAIntrinsic::GenISA_icmpxchgatomicraw:
             case GenISAIntrinsic::GenISA_fcmpxchgatomicraw:
-            case GenISAIntrinsic::GenISA_simdBlockRead:
+            case GenISAIntrinsic::GenISA_simdBlockRead: 
                 pBuffer = intr->getOperand(0);
                 break;
             case GenISAIntrinsic::GenISA_intatomicrawA64:
@@ -267,19 +267,19 @@ void PromoteResourceToDirectAS::PromoteSamplerTextureToDirectAS(GenIntrinsicInst
     BufferAccessType accTy;
     bool canPromote = false;
     Value* arrayIndex = nullptr;
-
+    
     std::vector<Value*> instList;
     Value* srcPtr = IGC::TracePointerSource(resourcePtr, false, true, instList);
-
+    
     if (srcPtr)
     {
         if (auto alloca = llvm::dyn_cast<AllocaInst>(srcPtr))
         {
             arrayIndex = FindArrayIndex(instList, builder);
             if (arrayIndex != nullptr)
             {
-                // TODO: We could read igc.read_only_array metadata attached to alloca.
-                // If not -1, it should contain base index of this array. In this case,
+                // TODO: We could read igc.read_only_array metadata attached to alloca. 
+                // If not -1, it should contain base index of this array. In this case, 
                 // FindArrayBaseArg would not be needed.
 
                 // Find input argument for the first element in this array.
@@ -289,7 +289,7 @@ void PromoteResourceToDirectAS::PromoteSamplerTextureToDirectAS(GenIntrinsicInst
     }
 
     if (srcPtr)
-    {
+    {   
         // Trace the resource pointer.
         // If we can find it, we can promote the indirect access to direct access
         // by encoding the BTI as a direct addrspace
@@ -308,7 +308,7 @@ void PromoteResourceToDirectAS::PromoteSamplerTextureToDirectAS(GenIntrinsicInst
             {
                 assert(m_pCodeGenContext->type == ShaderType::OPENCL_SHADER);
                 ModuleMetaData *modMD = getAnalysis<MetaDataUtilsWrapper>().getModuleMetaData();
-                if (modMD->FuncMD.find(function) != modMD->FuncMD.end())
+                if (modMD->FuncMD.find(function) != modMD->FuncMD.end()) 
                 {
                     FunctionMetaData *funcMD = &modMD->FuncMD[function];
                     ResourceAllocMD *resAllocMD = &funcMD->resAllocMD;
@@ -347,7 +347,7 @@ void PromoteResourceToDirectAS::PromoteSamplerTextureToDirectAS(GenIntrinsicInst
 
         addrSpace = IGC::EncodeAS4GFXResource(*bufferId, bufTy, 0);
         PointerType* newptrType = PointerType::get(resourcePtr->getType()->getPointerElementType(), addrSpace);
-
+        
         Value* mutePtr = nullptr;
         if (llvm::isa<llvm::ConstantInt>(bufferId))
         {
@@ -645,7 +645,7 @@ void PromoteResourceToDirectAS::PromoteBufferToDirectAS(Instruction* inst, Value
                 PointerType *ptrType = PointerType::get(pBufferType, directAS);
                 pBuffer = builder.CreateIntToPtr(offsetVal, ptrType);
 
-                unsigned alignment = (unsigned)llvm::cast<llvm::ConstantInt>(pIntr->getOperand(3))->getZExtValue();
+                unsigned alignment = pBufferType->getScalarSizeInBits() / 8;
 
                 // Promote storeraw back to store
                 Value* storeVal = pIntr->getOperand(2);
@@ -761,7 +761,7 @@ void PromoteResourceToDirectAS::GetAccessInstToSrcPointerMap(Instruction* inst,
         }
         else
             return;
-    }
+    }       
 
     Value* srcPtr = IGC::TracePointerSource(resourcePtr);
 

IGC/GenISAIntrinsics/GenIntrinsicInst.h

@@ -513,9 +513,6 @@ class StoreRawIntrinsic : public GenIntrinsicInst {
     static inline bool classof(const Value *V) {
         return isa<GenIntrinsicInst>(V) && classof(cast<GenIntrinsicInst>(V));
     }
-    inline Value* AlignmentValue() const {
-        return getOperand(3);
-    }
     inline Value* getOffsetValue() const
     {
         return getOperand(1);
@@ -527,11 +524,6 @@ class StoreRawIntrinsic : public GenIntrinsicInst {
     inline void setOffsetValue(Value* V) {
         setOperand(1, V);
     }
-    inline void setAlignment(unsigned int alignment)
-    {
-        Value* newAlignment = ConstantInt::get(getOperand(3)->getType(), alignment);
-        setOperand(3, newAlignment);
-    }
 };
 
 class AtomicRawIntrinsic : public GenIntrinsicInst {

IGC/GenISAIntrinsics/Intrinsic_definitions.py

@@ -87,8 +87,8 @@
     "GenISA_typedwrite": ["void",["anyptr","int","int","int","int","float","float","float","float"],"None"],
     "GenISA_ldraw_indexed": ["any:float",["anyptr","int", "int"],"ReadArgMem"],
     "GenISA_ldrawvector_indexed": ["anyvector",["anyptr","int", "int"],"ReadArgMem"],
-    "GenISA_storeraw_indexed": ["void",["anyptr","int","any:float", "int"],"None"],
-    "GenISA_storerawvector_indexed": ["void",["anyptr","int","anyvector", "int"],"None"],
+    "GenISA_storeraw_indexed": ["void",["anyptr","int","any:float"],"None"],
+    "GenISA_storerawvector_indexed": ["void",["anyptr","int","anyvector"],"None"],
     "GenISA_intatomicraw": ["anyint",["anyptr","int",0,"int"],"ReadWriteArgMem"],
     "GenISA_floatatomicraw": ["anyfloat",["anyptr","int",0,"int"],"ReadWriteArgMem"],
     "GenISA_intatomicrawA64": ["anyint",["anyptr","anyptr",0,"int"],"ReadWriteArgMem"],