Instruction combining is not able to combine circular phi-select

instructions leading to the same instruction.
Example:
Entry:
   %0 = load <4 x float>
While.body:
   %1 = phi <4 x float> [ %0, %entry ], [ %2, %while.end ]
   ...
   %2 = select i1 %switch_cond, <4 x float> %0, <4 x float> %1

Change-Id: I463d49a2dfae3a836e2c98d48b23096d3b3022c5

Author: aratajew

IGC/Compiler/CISACodeGen/ShaderCodeGen.cpp

@@ -1329,7 +1329,10 @@ void OptimizeIR(CodeGenContext* pContext)
             // Use CFGSimplification to do clean-up. Needs to be invoked before lowerSwitch.
             mpm.add(llvm::createCFGSimplificationPass());
 
-            if(IGC_IS_FLAG_DISABLED(DisableFlattenSmallSwitch))
+            // run instruction combining to clean up the code after CFG optimizations
+            mpm.add(createIGCInstructionCombiningPass());
+
+            if (IGC_IS_FLAG_DISABLED(DisableFlattenSmallSwitch))
             {
                 mpm.add(createFlattenSmallSwitchPass());
             }
@@ -1338,9 +1341,6 @@ void OptimizeIR(CodeGenContext* pContext)
             // After lowering 'switch', run jump threading to remove redundant jumps.
             mpm.add(llvm::createJumpThreadingPass());
 
-            // run instruction combining to clean up the code after CFG optimizations
-            mpm.add(createIGCInstructionCombiningPass());
-
             mpm.add(llvm::createDeadCodeEliminationPass());
             mpm.add(llvm::createEarlyCSEPass());