Turn on Regkey EnableWaveShuffleIndexSinking

Turn on WaveShuffleIndexSinking by default

Author: bowenxue-intel

IGC/Compiler/Optimizer/WaveShuffleIndexSinking.cpp

@@ -404,11 +404,15 @@ bool WaveShuffleIndexSinkingImpl::splitWaveShuffleIndexes()
         {
             if( auto* waveShuffleInst = dyn_cast<WaveShuffleIndexIntrinsic>( &I ) )
             {
-                if( !waveShuffleInst->getUniqueUndroppableUser() )
+                if( auto* constantChannel = dyn_cast<ConstantInt>( waveShuffleInst->getChannel() ) )
                 {
-                    // More than one user, split to potentially uncover more chances sink each individual WaveShuffleIndex
-                    Changed = true;
-                    InstsToSplit.push_back( waveShuffleInst );
+                    // Do not split WaveShuffleIndex insts that do not have a constant index since they cannot be optimized by this pass anyways
+                    if( !waveShuffleInst->getUniqueUndroppableUser() )
+                    {
+                        // More than one user, split to potentially uncover more chances sink each individual WaveShuffleIndex
+                        Changed = true;
+                        InstsToSplit.push_back( waveShuffleInst );
+                    }
                 }
             }
         }
@@ -423,7 +427,7 @@ bool WaveShuffleIndexSinkingImpl::splitWaveShuffleIndexes()
             auto* userInst = cast<Instruction>( user );
             auto* clonedWaveShuffleInst = instToSplit->clone();
             clonedWaveShuffleInst->setName( instToSplit->getName() + "_clone" );
-            clonedWaveShuffleInst->insertBefore( userInst );
+            clonedWaveShuffleInst->insertBefore( instToSplit );
             // Track replacement to perform after loop since iterators will be messed up if performed mid loop
             ReplacementPairs.emplace_back(userInst, clonedWaveShuffleInst );
         }

IGC/Compiler/tests/WaveShuffleIndexSinking/split-sink-merge.ll

@@ -29,22 +29,22 @@ define void @test_split_sink_merge(i32 %a, i32 %b, i32 %c, i32 %d) {
   %use1_ws0 = shl i32 %ws0, 2
 ; CHECK: add i32 [[WS0]], %c
   %anchor1_ws0 = add i32 %use1_ws0, %c
-; CHECK: [[WS1C1:%.*]] = call i32 @llvm.genx.GenISA.WaveShuffleIndex.i32(i32 [[USE1_WS0_WS1C1_HOISTED]], i32 1, i32 0)
+; CHECK: [[USE2_WS1C2_WS2C1_HOISTED:%.*]] = shl i32 %a, 3
+; CHECK-DAG: [[WS1C1:%.*]] = call i32 @llvm.genx.GenISA.WaveShuffleIndex.i32(i32 [[USE1_WS0_WS1C1_HOISTED]], i32 1, i32 0)
+; CHECK-DAG: [[WS1C2:%.*]] = call i32 @llvm.genx.GenISA.WaveShuffleIndex.i32(i32 [[USE2_WS1C2_WS2C1_HOISTED]], i32 1, i32 0)
   %ws1 = call i32 @llvm.genx.GenISA.WaveShuffleIndex.i32(i32 %a, i32 1, i32 0)
   %use1_ws1 = shl i32 %ws1, 2
 ; CHECK: add i32 [[WS1C1]], %c
   %anchor1_ws1 = add i32 %use1_ws1, %c
-; CHECK: [[USE2_WS1C2_WS2C1_HOISTED:%.*]] = shl i32 %a, 3
-; CHECK: [[WS1C2:%.*]] = call i32 @llvm.genx.GenISA.WaveShuffleIndex.i32(i32 [[USE2_WS1C2_WS2C1_HOISTED]], i32 1, i32 0)
   %use2_ws1 = shl i32 %ws1, 3
 ; CHECK: add i32 [[WS1C2]], %d
   %anchor2_ws1 = add i32 %use2_ws1, %d
-; CHECK: [[WS2C1:%.*]] = call i32 @llvm.genx.GenISA.WaveShuffleIndex.i32(i32 [[USE2_WS1C2_WS2C1_HOISTED]], i32 2, i32 0)
+; CHECK-DAG: [[WS2C1:%.*]] = call i32 @llvm.genx.GenISA.WaveShuffleIndex.i32(i32 [[USE2_WS1C2_WS2C1_HOISTED]], i32 2, i32 0)
+; CHECK-DAG: [[WS2C2:%.*]] = call i32 @llvm.genx.GenISA.WaveShuffleIndex.i32(i32 %a, i32 2, i32 0)
   %ws2 = call i32 @llvm.genx.GenISA.WaveShuffleIndex.i32(i32 %a, i32 2, i32 0)
   %use1_ws2 = shl i32 %ws2, 3
 ; CHECK: add i32 [[WS2C1]], %d
   %anchor1_ws2 = add i32 %use1_ws2, %d
-; CHECK: [[WS2C2:%.*]] = call i32 @llvm.genx.GenISA.WaveShuffleIndex.i32(i32 %a, i32 2, i32 0)
 ; CHECK: [[USE2_WS2C2_NOT_HOISTED:%.*]] = shl i32 [[WS2C2]], 4
   %use2_ws2 = shl i32 %ws2, 4
 ; CHECK: add i32 [[USE2_WS2C2_NOT_HOISTED]], %d

IGC/common/igc_flags.h

@@ -328,7 +328,7 @@ 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, EnableWaveShuffleIndexSinking, false, "Hoist identical instructions operating on WaveShuffleIndex instructions with the same source and a constant lane/channel", false)
+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.", true)