Revert "[IGC Performance][IGC]: Preprocess vector bitcasts to enable memory operations simplification."

This reverts commit a3b0a66933d6297c8aca4badc4e877558c4be7c2.

Author: eternastudento

IGC/Compiler/CISACodeGen/ShaderCodeGen.cpp

@@ -651,8 +651,6 @@ static void AddLegalizationPasses(CodeGenContext& ctx, IGCPassManager& mpm, PSSi
         mpm.add(createSplitIndirectEEtoSelPass());
     }
 
-    // Preprocess vector bitcasts to enable memory operations simplification.
-    mpm.add(createVectorBitCastOptPass());
     // Split big vector & 3-element load/store, etc.
     mpm.add(createVectorPreProcessPass());
 

IGC/Compiler/CustomSafeOptPass.cpp

@@ -473,6 +473,12 @@ void CustomSafeOptPass::visitCallInst(CallInst& C)
             break;
         }
 
+        case GenISAIntrinsic::GenISA_ldrawvector_indexed:
+        {
+            visitLdRawVec(inst);
+            break;
+        }
+
         default:
             break;
         }
@@ -1339,6 +1345,56 @@ void IGC::CustomSafeOptPass::visitLdptr(llvm::CallInst* inst)
     }
 }
 
+
+void IGC::CustomSafeOptPass::visitLdRawVec(llvm::CallInst* inst)
+{
+    //Try to optimize and remove vector ld raw and change to scalar ld raw
+
+    //%a = call <4 x float> @llvm.genx.GenISA.ldrawvector.indexed.v4f32.p1441792f32(
+    //.....float addrspace(1441792) * %243, i32 %offset, i32 4, i1 false), !dbg !216
+    //%b = extractelement <4 x float> % 245, i32 0, !dbg !216
+
+    //into
+
+    //%new_offset = add i32 %offset, 0, !dbg !216
+    //%b = call float @llvm.genx.GenISA.ldraw.indexed.f32.p1441792f32.i32.i32.i1(
+    //.....float addrspace(1441792) * %251, i32 %new_offset, i32 4, i1 false)
+
+    if (inst->hasOneUse() &&
+        isa<ExtractElementInst>(inst->user_back()))
+    {
+        auto EE = cast<ExtractElementInst>(inst->user_back());
+        if (auto constIndex = dyn_cast<ConstantInt>(EE->getIndexOperand()))
+        {
+            llvm::IRBuilder<> builder(inst);
+
+            llvm::SmallVector<llvm::Type*, 2> ovldtypes{
+                EE->getType(), //float type
+                inst->getOperand(0)->getType(),
+            };
+
+            // For new_offset we need to take into acount the index of the Extract
+            // and convert it to bytes and add it to the existing offset
+            auto new_offset = constIndex->getZExtValue() * 4;
+
+            llvm::SmallVector<llvm::Value*, 4> new_args{
+                inst->getOperand(0),
+                builder.CreateAdd(inst->getOperand(1),builder.getInt32((unsigned)new_offset)),
+                inst->getOperand(2),
+                inst->getOperand(3)
+            };
+
+            Function* pLdraw_indexed_intrinsic = llvm::GenISAIntrinsic::getDeclaration(
+                inst->getModule(),
+                GenISAIntrinsic::GenISA_ldraw_indexed,
+                ovldtypes);
+
+            llvm::Value* ldraw_indexed = builder.CreateCall(pLdraw_indexed_intrinsic, new_args, "");
+            EE->replaceAllUsesWith(ldraw_indexed);
+        }
+    }
+}
+
 void IGC::CustomSafeOptPass::visitSampleBptr(llvm::SampleIntrinsic* sampleInst)
 {
     // sampleB with bias_value==0 -> sample

IGC/Compiler/CustomSafeOptPass.hpp

@@ -80,6 +80,7 @@ namespace IGC
         void visitFPTruncInst(llvm::FPTruncInst& I);
         void visitExtractElementInst(llvm::ExtractElementInst& I);
         void visitLdptr(llvm::CallInst* inst);
+        void visitLdRawVec(llvm::CallInst* inst);
         void visitLoadInst(llvm::LoadInst& I);
         void dp4WithIdentityMatrix(llvm::ExtractElementInst& I);
         bool isIdentityMatrix(llvm::ExtractElementInst& I);