Match and fold reduction sequences into intrinsics in VC

Trying to match vector reduction sequences emitted by frontend compilers
and fold them into LLVM intrinsics. The intrinsics should be lowered by
the GenXLowering pass into efficient instruction sequences.

Author: vmustya

IGC/VectorCompiler/include/vc/GenXOpts/GenXOpts.h

@@ -1,6 +1,6 @@
 /*========================== begin_copyright_notice ============================
 
-Copyright (C) 2019-2023 Intel Corporation
+Copyright (C) 2019-2024 Intel Corporation
 
 SPDX-License-Identifier: MIT
 
@@ -91,6 +91,9 @@ void initializeGenXTranslateIntrinsicsPass(PassRegistry &);
 FunctionPass *createGenXTypeLegalizationPass();
 void initializeGenXTypeLegalizationPass(PassRegistry &);
 
+FunctionPass *createGenXFoldReductionPass();
+void initializeGenXFoldReductionPass(PassRegistry &);
+
 } // End llvm namespace
 
 #endif // VC_GENXOPTS_GENXOPTS_H

IGC/VectorCompiler/lib/GenXCodeGen/GenXTargetMachine.cpp

@@ -208,6 +208,7 @@ void initializeGenXPasses(PassRegistry &registry) {
   initializeGenXLegacyToLscTranslatorPass(registry);
   initializeGenXSLMResolutionPass(registry);
   initializeGenXTypeLegalizationPass(registry);
+  initializeGenXFoldReductionPass(registry);
   // WRITE HERE MORE PASSES IF IT'S NEEDED;
 }
 
@@ -869,6 +870,7 @@ void GenXTargetMachine::adjustPassManager(PassManagerBuilder &PMBuilder) {
     PM.add(createGenXTranslateSPIRVBuiltinsPass());
     PM.add(createAlwaysInlinerLegacyPass());
     PM.add(createAlwaysInlinerLegacyPass());
+    PM.add(createGenXFoldReductionPass());
     PM.add(createGenXPrintfResolutionPass());
     PM.add(createGenXImportOCLBiFPass());
     PM.add(createGenXBIFFlagCtrlResolutionPass());