[GPU] Fix hoisting after upstream change disallowing view ops (iree-o…

…rg#15192)

MLIR changed to disallow hoisting involving memref view-like ops, which
is heavily relied on in LLVMGPU and SPIR-V CodeGen for better
performance. Now we need to first fold aliased memrefs and then perform
hoisting.

This drops the local revert for llvm/llvm-project@94c04772
It additionally cherry-picks the following commits:
* llvm/llvm-project@ebaf8d49
* llvm/llvm-project@3049ac44

Fixes iree-org#15083

compiler/src/iree/compiler/Codegen/LLVMGPU/Passes.cpp

@@ -239,7 +239,8 @@ void addGPUMatmulTensorCorePassPipeline(OpPassManager &pm,
   // Linalg -> vector
   nestedModulePM.addNestedPass<func::FuncOp>(
       createLLVMGPUTensorCoreVectorizationPass());
-  nestedModulePM.addNestedPass<func::FuncOp>(createCanonicalizerPass());
+  nestedModulePM.addNestedPass<func::FuncOp>(
+      memref::createFoldMemRefAliasOpsPass());
   nestedModulePM.addNestedPass<func::FuncOp>(createCSEPass());
   nestedModulePM.addNestedPass<func::FuncOp>(
       createOptimizeVectorTransferPass());
@@ -302,7 +303,8 @@ void addGPUMatmulTensorCoreMmaSyncPassPipeline(OpPassManager &pm,
   // Linalg -> vector
   nestedModulePM.addNestedPass<func::FuncOp>(
       createLLVMGPUTensorCoreVectorizationPass(GPUTensorCoreType::MMA_SYNC));
-  nestedModulePM.addNestedPass<func::FuncOp>(createCanonicalizerPass());
+  nestedModulePM.addNestedPass<func::FuncOp>(
+      memref::createFoldMemRefAliasOpsPass());
   nestedModulePM.addNestedPass<func::FuncOp>(createCSEPass());
   nestedModulePM.addNestedPass<func::FuncOp>(
       createOptimizeVectorTransferPass());

compiler/src/iree/compiler/Codegen/LLVMGPU/test/conv_pipeline_test.mlir

@@ -82,17 +82,17 @@ module attributes {hal.device.targets = [#device_target_cuda]} {
 //   CHECK-LABEL:  func.func @conv_nchw
 // TODO: hoist the accumulator read and fold the transfer_write.
 //         CHECK:    vector.transfer_write
+// CHECK-COUNT-4:    vector.transfer_read
 //         CHECK:    scf.for
 //         CHECK:      scf.for
-// CHECK-COUNT-3:        vector.transfer_read
-//         CHECK:        vector.contract
-//         CHECK:        vector.transfer_write
 // CHECK-COUNT-2:        vector.transfer_read
 //         CHECK:        vector.contract
-//         CHECK:        vector.transfer_write
-// CHECK-COUNT-2:        vector.transfer_read
+//         CHECK:        vector.transfer_read
 //         CHECK:        vector.contract
-//         CHECK:        vector.transfer_write
-// CHECK-COUNT-2:        vector.transfer_read
+//         CHECK:        vector.transfer_read
+//         CHECK:        vector.contract
+//         CHECK:        vector.transfer_read
 //         CHECK:        vector.contract
-//         CHECK:        vector.transfer_write
+//         CHECK:      scf.yield
+//         CHECK:    scf.yield
+// CHECK-COUNT-4:    vector.transfer_write

compiler/src/iree/compiler/Codegen/SPIRV/Passes.cpp

@@ -374,9 +374,15 @@ void addSPIRVCooperativeMatrixVectorizePassPipeline(OpPassManager &pm,
         createGenericVectorizationPass(options));
   }
 
+  // With subview ops, vector hoisting won't kick in. So fold memref subview ops
+  // before performing vector unrolling and hoisting.
+  nestedModulePM.addNestedPass<func::FuncOp>(
+      memref::createFoldMemRefAliasOpsPass());
+
   // Vectorize to cooperative ops.
   nestedModulePM.addNestedPass<func::FuncOp>(
       createSPIRVVectorizeToCooperativeOpsPass());
+  nestedModulePM.addPass(createCSEPass());
   nestedModulePM.addNestedPass<func::FuncOp>(
       createHoistRedundantVectorTransfersPass());
   nestedModulePM.addNestedPass<func::FuncOp>(
@@ -392,10 +398,6 @@ void addSPIRVCooperativeMatrixVectorizePassPipeline(OpPassManager &pm,
   nestedModulePM.addNestedPass<func::FuncOp>(createOptimizeVectorTransferPass(
       /*flatten=*/false, /*dropUnitDims=*/false));
 
-  // Fold subview ops is reqiured for converting vector transfer ops into SPIR-V
-  // cooperative ops in the next step.
-  nestedModulePM.addPass(memref::createFoldMemRefAliasOpsPass());
-
   nestedModulePM.addNestedPass<func::FuncOp>(createForOpCanonicalizationPass());
   nestedModulePM.addPass(createCanonicalizerPass());
   nestedModulePM.addPass(createCSEPass());
@@ -466,19 +468,19 @@ void addSPIRVMatmulPromoteVectorizePassPipeline(OpPassManager &topPM,
   nestedPM.addNestedPass<func::FuncOp>(createGPUReduceSharedMemoryBankConflicts(
       detail::bankConflictReductionPaddingBits));
 
-  addSPIRVVectorLoweringPasses(nestedPM);
+  // With subview ops, vector hoisting won't kick in. So fold memref subview ops
+  // before performing vector unrolling and hoisting.
+  nestedPM.addNestedPass<func::FuncOp>(memref::createFoldMemRefAliasOpsPass());
+
+  nestedPM.addNestedPass<func::FuncOp>(createSPIRVInitialVectorLoweringPass());
+  nestedPM.addPass(createCSEPass());
+  nestedPM.addNestedPass<func::FuncOp>(
+      createHoistRedundantVectorTransfersPass());
+  nestedPM.addNestedPass<func::FuncOp>(createSPIRVFinalVectorLoweringPass());
+
   nestedPM.addNestedPass<func::FuncOp>(createForOpCanonicalizationPass());
   nestedPM.addPass(createCanonicalizerPass());
   nestedPM.addPass(createCSEPass());
-  // After vectorization and some basic cleanup, optimize vector transfer ops.
-  // Here we won't have large n-D vectors being put as loop carried values due
-  // to hoisting. Because this is before folding all memref subview ops away, we
-  // still have subview ops using the same indices, which allows for transfer
-  // read/write forwarding.
-  nestedPM.addNestedPass<func::FuncOp>(createOptimizeVectorTransferPass(
-      /*flatten=*/false, /*dropUnitDims=*/false));
-
-  nestedPM.addNestedPass<func::FuncOp>(memref::createFoldMemRefAliasOpsPass());
   nestedPM.addNestedPass<func::FuncOp>(createOptimizeVectorTransferPass(
       /*flatten=*/false, /*dropUnitDims=*/false));
 

compiler/src/iree/compiler/Codegen/TransformStrategies/GPU/Common.cpp

@@ -570,6 +570,11 @@ Value mlir::iree_compiler::gpu::buildConvertToTensorCoreOp(
   // be replaced by a single transform.
   b.create<SynchronizeLoopOp>(forH);
 
+  b.create<transform::ApplyPatternsOp>(funcH, [](OpBuilder &b, Location loc) {
+    b.create<transform::ApplyFoldMemrefAliasOpsPatternsOp>(loc);
+  });
+  b.create<IREE::transform_dialect::ApplyCommonSubexpressionEliminationOp>(
+      funcH);
   // TODO: not a functional style transform and avoid returning funcH.
   funcH = b.create<transform::HoistRedundantVectorTransfersOp>(
       transform::AnyOpType::get(b.getContext()), funcH);