Do not consider subgroup-uniform values (e.g., subgroup intrinsics and dispatch mask) as uniform for certain platforms.

Change-Id: I95a6340b0a950870f0bd33d28fc4932368783efd

Author: weiyu-chen

IGC/Compiler/CISACodeGen/Platform.hpp

@@ -382,6 +382,12 @@ namespace IGC
     {
         return IGC_IS_FLAG_ENABLED(FuseTypedWrite);
     }
+
+    bool HasUniformSubGroupIntrinsic() const
+    {
+        return IGC_IS_FLAG_DISABLED(NonUniformSubGroupIntrinsic) &&
+            m_platformInfo.eProductFamily != IGFX_TIGERLAKE_LP;
+    }
     // ***** Below go accessor methods for testing WA data from WA_TABLE *****
 
     bool WaDoNotPushConstantsForAllPulledGSTopologies() const

IGC/Compiler/CISACodeGen/WIAnalysis.cpp

@@ -1131,32 +1131,45 @@ WIAnalysis::WIDependancy WIAnalysisRunner::calculate_dep(const CallInst* inst)
             return WIAnalysis::UNIFORM;
         }
 
-        if (intrinsic_name == llvm_waveShuffleIndex)
+        if (m_CGCtx->platform.HasUniformSubGroupIntrinsic())
         {
-            Value* op0 = inst->getArgOperand(0);
-            Value* op1 = inst->getArgOperand(1);
-            if (WIAnalysis::UNIFORM == getDependency(op0) ||
-                WIAnalysis::UNIFORM == getDependency(op1))
+            if (intrinsic_name == llvm_waveShuffleIndex)
+            {
+                Value* op0 = inst->getArgOperand(0);
+                Value* op1 = inst->getArgOperand(1);
+                if (WIAnalysis::UNIFORM == getDependency(op0) ||
+                    WIAnalysis::UNIFORM == getDependency(op1))
+                {
+                    return WIAnalysis::UNIFORM;
+                }
+            }
+
+            if (intrinsic_name == llvm_waveBallot || intrinsic_name == llvm_waveAll)
             {
                 return WIAnalysis::UNIFORM;
             }
-        }
 
-        if (intrinsic_name == llvm_waveBallot || intrinsic_name == llvm_waveAll)
-        {
-            return WIAnalysis::UNIFORM;
-        }
+            if (intrinsic_name == llvm_waveClustered)
+            {
+                const unsigned clusterSize = static_cast<unsigned>(
+                    cast<llvm::ConstantInt>(inst->getArgOperand(2))->getZExtValue());
 
-        if (intrinsic_name == llvm_waveClustered)
+                constexpr unsigned maxSimdSize = 32;
+                if (clusterSize == maxSimdSize)
+                {
+                    // TODO: do the same for SIMD8 and SIMD16 if possible.
+                    return WIAnalysis::UNIFORM;
+                }
+            }
+        }
+        else
         {
-            const unsigned clusterSize = static_cast<unsigned>(
-                cast<llvm::ConstantInt>(inst->getArgOperand(2))->getZExtValue());
-
-            constexpr unsigned maxSimdSize = 32;
-            if (clusterSize == maxSimdSize)
+            if (GII_id == GenISAIntrinsic::GenISA_getSR0 ||
+                GII_id == GenISAIntrinsic::GenISA_eu_id ||
+                GII_id == GenISAIntrinsic::GenISA_eu_thread_id)
             {
-                // TODO: do the same for SIMD8 and SIMD16 if possible.
-                return WIAnalysis::UNIFORM;
+                // do not assume these per-thread values are globally uniform
+                return WIAnalysis::RANDOM;
             }
         }
 

IGC/common/igc_flags.def

@@ -282,6 +282,7 @@ DECLARE_IGC_REGKEY(bool, UniformMemOptLimit,            0,     "Limit of uniform
 DECLARE_IGC_REGKEY(bool, EnableFunctionPointer,         true,  "Enables support for function pointers and indirect calls", false)
 DECLARE_IGC_REGKEY(bool, EnableIndirectCallOptimization, false, "Enables inlining indirect calls by comparing function addresses", false)
 DECLARE_IGC_REGKEY(bool, ForceFFIDOverwrite,            false,   "Force overwriting ffid in sr0.0", false)
+DECLARE_IGC_REGKEY(bool, NonUniformSubGroupIntrinsic,   false, "Consider sub-group intrinsics to be non-uniform", false)
 
 DECLARE_IGC_REGKEY(bool, EnableReadGTPinInput,          true,  "Enables setting GTPin context flags by reading the input to the compiler adapters", false)