This change is needed for it is risky to mess with precision in vertex shaders by contracting fmul+(add/sub) into a mad going into a trig function like sin,

otherwise the position values can become off their expected values.

Change-Id: I5e45ae6fc2bd6110a87113501e3248181de32d23

Author: paigeale

IGC/Compiler/CISACodeGen/PatternMatchPass.cpp

@@ -1989,6 +1989,21 @@ namespace IGC
         {
             return false;
         }
+
+        using namespace llvm::PatternMatch;
+        Value* LHS = NULL, * RHS = NULL, * Op1 = NULL, * Op0 = NULL;
+        if (m_ctx->type == ShaderType::VERTEX_SHADER &&
+            m_ctx->m_DriverInfo.PreventZFighting() &&
+            (match(&I,m_BinOp(m_FMul(m_Value(LHS), m_Value(RHS)), m_Value(Op1))) ||
+             match(&I,m_BinOp(m_Value(Op0), m_FMul(m_Value(LHS), m_Value(RHS))))) &&
+            I.hasOneUse() && isa<IntrinsicInst>(I.user_back()))
+        {
+            auto fmul_val = Op0 != NULL ? I.getOperand(1) : I.getOperand(0);
+            auto fmul = cast<BinaryOperator>(fmul_val);
+            if (cast<IntrinsicInst>(I.user_back())->getIntrinsicID() == llvm::Intrinsic::sin &&
+                m_PosDep->PositionDependsOnInst(fmul) && NeedInstruction(*fmul))
+                return false;
+        }
         if (IGC_IS_FLAG_ENABLED(DisableMatchMad))
         {
             return false;