[SYCL-MLIR]: Improve Polygeist LICM to handle invariant operations in scf::ForOp and AffineForOp (#8377)

This PR improves the `ParallelLICM` pass in polygeist as follows:
- adds support for non-parallel loop constructs scf::ForOp and improves
handleing of AffineForOp
- fixes a problem, in sequential loops function calls are conservatively
considered to modify memory
- add more tests for both reduction loops and loops that do not yield a
value
   - renames `ParallelLICM` pass to simply `LICM`

---------

Signed-off-by: Tiotto, Ettore <[email protected]>

Author: etiotto

polygeist/include/polygeist/Passes/Passes.h

@@ -9,7 +9,7 @@ class PatternRewriter;
 class DominanceInfo;
 class LLVMTypeConverter;
 namespace polygeist {
-std::unique_ptr<Pass> createParallelLICMPass();
+std::unique_ptr<Pass> createLICMPass();
 std::unique_ptr<Pass> createMem2RegPass();
 std::unique_ptr<Pass> createLoopRestructurePass();
 std::unique_ptr<Pass> createInnerSerializationPass();

polygeist/include/polygeist/Passes/Passes.td

@@ -65,9 +65,9 @@ def SCFCanonicalizeFor : Pass<"canonicalize-scf-for"> {
   let constructor = "mlir::polygeist::createCanonicalizeForPass()";
 }
 
-def ParallelLICM : Pass<"parallel-licm"> {
+def LICM : Pass<"licm"> {
   let summary = "Perform LICM on known parallel (and serial) loops";
-  let constructor = "mlir::polygeist::createParallelLICMPass()";
+  let constructor = "mlir::polygeist::createLICMPass()";
 
   let statistics = [
     Statistic<"numOpHoisted", "num-operations-hoisted", "Number of operations hoisted">

polygeist/lib/polygeist/Passes/CMakeLists.txt

@@ -6,10 +6,10 @@ add_mlir_dialect_library(MLIRPolygeistTransforms
   ConvertPolygeistToLLVM.cpp  
   InnerSerialization.cpp  
   LegalizeForSPIRV.cpp  
+  LICM.cpp
   LoopRestructure.cpp
   Mem2Reg.cpp
   OpenMPOpt.cpp
-  ParallelLICM.cpp
   ParallelLoopDistribute.cpp
   ParallelLower.cpp
   RaiseToAffine.cpp