[SYCL] Add barrier optimization pass (#19353)

It removes redundant barriers (both back-to-back and in general in CFG)
and downgrades global barrier to local if there are no global memory
accesses 'between' them. See description in
SYCLOptimizeBackToBackBarrier.cpp for more details.

---------

Signed-off-by: Sidorov, Dmitry <dmitry.sidorov@intel.com>
Signed-off-by: Dmitry Sidorov <dmitrii.s.sidorov@gmail.com>

Author: MrSidims

clang/lib/CodeGen/BackendUtil.cpp

@@ -52,6 +52,7 @@
 #include "llvm/SYCLLowerIR/SYCLAddOptLevelAttribute.h"
 #include "llvm/SYCLLowerIR/SYCLConditionalCallOnDevice.h"
 #include "llvm/SYCLLowerIR/SYCLCreateNVVMAnnotations.h"
+#include "llvm/SYCLLowerIR/SYCLOptimizeBarriers.h"
 #include "llvm/SYCLLowerIR/SYCLPropagateAspectsUsage.h"
 #include "llvm/SYCLLowerIR/SYCLPropagateJointMatrixUsage.h"
 #include "llvm/SYCLLowerIR/SYCLVirtualFunctionsAnalysis.h"
@@ -1080,13 +1081,16 @@ void EmitAssemblyHelper::RunOptimizationPipeline(
             MPM.addPass(ESIMDRemoveHostCodePass());
           });
 
-    // Add the InferAddressSpaces pass for all the SPIR[V] targets
+    // Add the InferAddressSpaces and SYCLOptimizeBarriers passes for all
+    // the SPIR[V] targets
     if (TargetTriple.isSPIR() || TargetTriple.isSPIRV()) {
       PB.registerOptimizerLastEPCallback(
           [](ModulePassManager &MPM, OptimizationLevel Level,
              ThinOrFullLTOPhase) {
             MPM.addPass(createModuleToFunctionPassAdaptor(
                 InferAddressSpacesPass(clang::targets::SPIR_GENERIC_AS)));
+            MPM.addPass(
+                createModuleToFunctionPassAdaptor(SYCLOptimizeBarriersPass()));
           });
     }
 

…LLowerIR/SYCLOptimizeBackToBackBarrier.h …/llvm/SYCLLowerIR/SYCLOptimizeBarriers.hllvm/include/llvm/SYCLLowerIR/SYCLOptimizeBackToBackBarrier.h renamed to llvm/include/llvm/SYCLLowerIR/SYCLOptimizeBarriers.h llvm/include/llvm/SYCLLowerIR/SYCLOptimizeBackToBackBarrier.h renamed to llvm/include/llvm/SYCLLowerIR/SYCLOptimizeBarriers.h

@@ -1,29 +1,29 @@
-//==- SYCLOptimizeBackToBackBarrier.h - SYCLOptimizeBackToBackBarrier Pass -==//
+//==- SYCLOptimizeBarriers.h - SYCLOptimizeBarriers Pass -==//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
-// This pass cleans up back-to-back ControlBarrier calls.
+// This pass cleans up ControlBarrier and MemoryBarrier calls.
 //
 //===----------------------------------------------------------------------===//
-#ifndef LLVM_SYCL_OPTIMIZE_BACK_TO_BACK_BARRIER_H
-#define LLVM_SYCL_OPTIMIZE_BACK_TO_BACK_BARRIER_H
+#ifndef LLVM_SYCL_OPTIMIZE_BARRIERS_H
+#define LLVM_SYCL_OPTIMIZE_BARRIERS_H
 
 #include "llvm/IR/PassManager.h"
 
 namespace llvm {
 
-class SYCLOptimizeBackToBackBarrierPass
-    : public PassInfoMixin<SYCLOptimizeBackToBackBarrierPass> {
+class SYCLOptimizeBarriersPass
+    : public PassInfoMixin<SYCLOptimizeBarriersPass> {
 public:
-  PreservedAnalyses run(Module &M, ModuleAnalysisManager &);
+  PreservedAnalyses run(Function &F, FunctionAnalysisManager &);
 
   static bool isRequired() { return true; }
 };
 
 } // namespace llvm
 
-#endif // LLVM_SYCL_OPTIMIZE_BACK_TO_BACK_BARRIER_H
+#endif // LLVM_SYCL_OPTIMIZE_BARRIERS_H

llvm/lib/Passes/PassBuilder.cpp

@@ -192,7 +192,7 @@
 #include "llvm/SYCLLowerIR/SYCLConditionalCallOnDevice.h"
 #include "llvm/SYCLLowerIR/SYCLCreateNVVMAnnotations.h"
 #include "llvm/SYCLLowerIR/SYCLJointMatrixTransform.h"
-#include "llvm/SYCLLowerIR/SYCLOptimizeBackToBackBarrier.h"
+#include "llvm/SYCLLowerIR/SYCLOptimizeBarriers.h"
 #include "llvm/SYCLLowerIR/SYCLPropagateAspectsUsage.h"
 #include "llvm/SYCLLowerIR/SYCLPropagateJointMatrixUsage.h"
 #include "llvm/SYCLLowerIR/SYCLVirtualFunctionsAnalysis.h"

llvm/lib/Passes/PassRegistry.def

@@ -180,7 +180,6 @@ MODULE_PASS("esimd-remove-host-code", ESIMDRemoveHostCodePass());
 MODULE_PASS("esimd-remove-optnone-noinline", ESIMDRemoveOptnoneNoinlinePass());
 MODULE_PASS("sycl-conditional-call-on-device", SYCLConditionalCallOnDevicePass())
 MODULE_PASS("sycl-joint-matrix-transform", SYCLJointMatrixTransformPass())
-MODULE_PASS("sycl-optimize-back-to-back-barrier", SYCLOptimizeBackToBackBarrierPass())
 MODULE_PASS("sycl-propagate-aspects-usage", SYCLPropagateAspectsUsagePass())
 MODULE_PASS("sycl-propagate-joint-matrix-usage", SYCLPropagateJointMatrixUsagePass())
 MODULE_PASS("sycl-add-opt-level-attribute", SYCLAddOptLevelAttributePass())
@@ -507,6 +506,7 @@ FUNCTION_PASS("slp-vectorizer", SLPVectorizerPass())
 FUNCTION_PASS("slsr", StraightLineStrengthReducePass())
 FUNCTION_PASS("stack-protector", StackProtectorPass(TM))
 FUNCTION_PASS("strip-gc-relocates", StripGCRelocates())
+FUNCTION_PASS("sycl-optimize-barriers", SYCLOptimizeBarriersPass())
 FUNCTION_PASS("tailcallelim", TailCallElimPass())
 FUNCTION_PASS("transform-warning", WarnMissedTransformationsPass())
 FUNCTION_PASS("trigger-crash-function", TriggerCrashFunctionPass())

llvm/lib/SYCLLowerIR/CMakeLists.txt

@@ -74,7 +74,7 @@ add_llvm_component_library(LLVMSYCLLowerIR
   SYCLDeviceRequirements.cpp
   SYCLKernelParamOptInfo.cpp
   SYCLJointMatrixTransform.cpp
-  SYCLOptimizeBackToBackBarrier.cpp
+  SYCLOptimizeBarriers.cpp
   SYCLPropagateAspectsUsage.cpp
   SYCLPropagateJointMatrixUsage.cpp
   SYCLVirtualFunctionsAnalysis.cpp