Turn on EnableUniformSimd1Stores pass for CS

Turn on EnableUniformSimd1Stores to use SIMD1 for uniform SLM Stores

Author: iwwu

IGC/Compiler/CISACodeGen/DriverInfo.hpp

@@ -383,6 +383,7 @@ namespace IGC
         virtual bool supportLscSamplerRouting() const { return true; }
         virtual bool supportBarrierControlFlowOptimization() const { return false; }
         virtual bool getLscStoresWithNonDefaultL1CacheControls() const { return true; }
+        virtual bool supportMergeUniformSimd1Stores() const { return true; }
 
         // Informs if the UMD understands atomic pull tile walk for raytracing
         virtual bool supportsAtomicPullSWTileWalk() const { return false; }

IGC/Compiler/CISACodeGen/Platform.hpp

@@ -1930,5 +1930,10 @@ bool allowsMoviForType(VISA_Type type) const {
   return (type == ISA_TYPE_UD || type == ISA_TYPE_D);
 }
 
+bool supportMergeUniformSimd1Stores() const
+{
+    return isCoreChildOf(IGFX_XE2_HPG_CORE);
+}
+
 };
 }//namespace IGC

IGC/Compiler/CISACodeGen/ShaderCodeGen.cpp

@@ -1919,6 +1919,8 @@ void OptimizeIR(CodeGenContext* const pContext)
         }
 
         if (pContext->type == ShaderType::COMPUTE_SHADER &&
+            pContext->m_DriverInfo.supportMergeUniformSimd1Stores() &&
+            pContext->platform.supportMergeUniformSimd1Stores() &&
             IGC_IS_FLAG_ENABLED(EnableUniformSimd1Stores))
         {
             mpm.add(new MergeUniformStores());

IGC/common/igc_flags.h

@@ -341,7 +341,7 @@ DECLARE_IGC_REGKEY(DWORD, FPRoundingModeCoalescingMaxDistance, 20, "Max distance
 DECLARE_IGC_REGKEY(bool, DisableDotAddToDp4aMerge, false, "Disable Dot and Add ops to Dp4a merge optimization.", false)
 DECLARE_IGC_REGKEY(bool, DisableLoopSplitWidePHIs, false, "Disable splitting of loop PHI values to eliminate subvector extract operations", false)
 DECLARE_IGC_REGKEY(bool, EnableBarrierControlFlowOptimizationPass, false, "Enable barrier control flow optimization pass", false)
-DECLARE_IGC_REGKEY(bool, EnableUniformSimd1Stores, false, "Collapse SIMD32 uniform stores into SIMD1 to reduce pressure", true)
+DECLARE_IGC_REGKEY(bool, EnableUniformSimd1Stores, true, "Collapse SIMD32 uniform stores into SIMD1 to reduce pressure", true)
 DECLARE_IGC_REGKEY(bool, EnableWaveShuffleIndexSinking, true, "Hoist identical instructions operating on WaveShuffleIndex instructions with the same source and a constant lane/channel", false)
 DECLARE_IGC_REGKEY(DWORD, WaveShuffleIndexSinkingMaxIterations, 3, "Max number of iterations to run iterative WaveShuffleIndexSinking", false)
 DECLARE_IGC_REGKEY(bool, EnableWaveAllJointReduction, false, "Enable Joint Reduction Optimization.", false)