Fixed the alignment for indirect float/half dst

Float/half operations requires all operands to have the same offsets.
If it is indirect, we have to assume they are not aligned so that
raw mov insts are generated. Raw mov instructions can be changed to
int type, as int indirect addressing is more flexible.

This change make sure that fixDstAlignment() will be invoked for
indirect float/half dst.

Author: jgu222

visa/HWConformity.cpp

@@ -1746,6 +1746,29 @@ bool HWConformity::fixDstAlignment(INST_LIST_ITER i, G4_BB* bb, G4_Type extype,
         return false;
     }
 
+    auto allScalarSrcs = [](G4_INST* I)
+    {
+        // return true If all of I's srcs are scalar; false otherwise.
+        bool allScalarSrc = true;
+        for (int i = 0, nsrc = I->getNumSrc(); i < nsrc; i++)
+        {
+            G4_Operand* src = I->getSrc(i);
+            if (src && !src->isNullReg() && src->isSrcRegRegion() &&
+                !(src->asSrcRegRegion()->getRegion()->isScalar()))
+            {
+                allScalarSrc = false;
+                break;
+            }
+        }
+        return allScalarSrc;
+    };
+
+    // float operation requires that operands of every channel to start on the same offset
+    // except for scalar operands.
+    // If dst is indirect and not all srcs are scalar, assume it is not aligned.
+    bool FPDstUnaligned_restriction9_7 =
+        ((extype == Type_F || extype == Type_HF) && dst->getRegAccess() != Direct && !allScalarSrcs(inst));
+
     bool dstHFMixModeInst = inst->getDst()->getType() == builder.getMixModeType() && extype == Type_F;
     bool dstNotAlignedToExecType = exec_size > 1 && (dst_elsize * h_stride) < extypesize &&
         !(builder.hasMixMode() && dstHFMixModeInst);
@@ -1764,7 +1787,8 @@ bool HWConformity::fixDstAlignment(INST_LIST_ITER i, G4_BB* bb, G4_Type extype,
             dst->getRegAccess() != Direct &&
             !(byteDst && extypesize == 2 && exec_size == 1)
            ) ||
-        dstNotAlignedToExecType)
+        dstNotAlignedToExecType ||
+        FPDstUnaligned_restriction9_7)
     {
         /*
          * 10.3

visa/ReduceExecSize.cpp

@@ -164,7 +164,11 @@ bool HWConformity::fixInstOpndTypeAlign(INST_LIST_ITER i, G4_BB* bb)
         if (extypesize < (int)kernel.numEltPerGRF<Type_UB>()/2)
         {
             uint32_t dst_elsize = inst->getDst()->getTypeSize();
-            if (dst_elsize < (unsigned int)extypesize)
+            if (dst_elsize < (unsigned int)extypesize ||
+                // indirect float type needs to be handled as well.
+                // See fixDstAlignment for detail
+                ((extype == Type_F || extype == Type_HF) &&
+                 inst->getDst()->getRegAccess() != Direct))
             {
                 if (fixDstAlignment(i, bb, extype, dst_elsize))
                 {