new LLVM optimization pass list

- use InstSimplify instead of InstCombine in some cases to speed it up
- reorder some passes
- add LoopLoadElimination and DivRemPairs

src/jitlayers.cpp

@@ -33,6 +33,7 @@
 #include <llvm/Transforms/Instrumentation.h>
 #include <llvm/Transforms/Vectorize.h>
 #include <llvm/Transforms/Scalar/GVN.h>
+#include <llvm/Transforms/Scalar/InstSimplifyPass.h>
 #include <llvm/Transforms/IPO/AlwaysInliner.h>
 #include <llvm/Transforms/InstCombine/InstCombine.h>
 #include <llvm/Support/SmallVectorMemoryBuffer.h>
@@ -122,13 +123,15 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level,
  PM->add(llvm::createMemorySanitizerPass(true));
 #endif
  if (opt_level < 2) {
- PM->add(createCFGSimplificationPass()); // Clean up disgusting code
+ PM->add(createCFGSimplificationPass());
  if (opt_level == 1) {
- PM->add(createSROAPass()); // Break up aggregate allocas
- PM->add(createInstructionCombiningPass()); // Cleanup for scalarrepl.
+ PM->add(createSROAPass());
+ PM->add(createInstructionCombiningPass());
  PM->add(createEarlyCSEPass());
+ // maybe add GVN?
+ // also try GVNHoist and GVNSink
  }
- PM->add(createMemCpyOptPass()); // Remove memcpy / form memset
+ PM->add(createMemCpyOptPass());
  PM->add(createAlwaysInlinerLegacyPass()); // Respect always_inline
  if (lower_intrinsics) {
  PM->add(createBarrierNoopPass());
@@ -138,7 +141,7 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level,
  PM->add(createFinalLowerGCPass());
  PM->add(createLowerPTLSPass(dump_native));
  }
- PM->add(createLowerSimdLoopPass());  // Annotate loop marked with "loopinfo" as LLVM parallel loop
+ PM->add(createLowerSimdLoopPass()); // Annotate loop marked with "loopinfo" as LLVM parallel loop
  if (dump_native)
  PM->add(createMultiVersioningPass());
  return;
@@ -149,43 +152,38 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level,
  if (opt_level >= 3) {
  PM->add(createBasicAAWrapperPass());
  }
- // list of passes from vmkit
- PM->add(createCFGSimplificationPass()); // Clean up disgusting code
+
+ PM->add(createCFGSimplificationPass());
  PM->add(createDeadCodeEliminationPass());
- PM->add(createSROAPass()); // Kill useless allocas
+ PM->add(createSROAPass());
 
- PM->add(createMemCpyOptPass());
+ //PM->add(createMemCpyOptPass());
 
  PM->add(createAlwaysInlinerLegacyPass()); // Respect always_inline
 
  // Running `memcpyopt` between this and `sroa` seems to give `sroa` a hard time
  // merging the `alloca` for the unboxed data and the `alloca` created by the `alloc_opt`
  // pass.
  PM->add(createAllocOptPass());
- PM->add(createInstructionCombiningPass()); // Cleanup for scalarrepl.
- // Now that SROA has cleaned up for front-end mess, a lot of control flow should
- // be more evident - try to clean it up.
- PM->add(createCFGSimplificationPass()); // Merge & remove BBs
+ // consider AggressiveInstCombinePass at optlevel > 2
+ PM->add(createInstructionCombiningPass());
+ PM->add(createCFGSimplificationPass());
  if (dump_native)
  PM->add(createMultiVersioningPass());
- PM->add(createSROAPass()); // Break up aggregate allocas
- PM->add(createInstructionCombiningPass()); // Cleanup for scalarrepl.
- PM->add(createJumpThreadingPass()); // Thread jumps.
- PM->add(createInstructionCombiningPass()); // Combine silly seq's
-
- //PM->add(createCFGSimplificationPass()); // Merge & remove BBs
- PM->add(createReassociatePass()); // Reassociate expressions
+ PM->add(createSROAPass());
+ PM->add(createInstSimplifyLegacyPass());
+ PM->add(createJumpThreadingPass());
 
- // this has the potential to make some things a bit slower
- //PM->add(createBBVectorizePass());
+ PM->add(createReassociatePass());
 
- PM->add(createEarlyCSEPass()); //// ****
+ PM->add(createEarlyCSEPass());
 
  // Load forwarding above can expose allocations that aren't actually used
  // remove those before optimizing loops.
  PM->add(createAllocOptPass());
- PM->add(createLoopIdiomPass()); //// ****
- PM->add(createLoopRotatePass()); // Rotate loops.
+ PM->add(createLoopRotatePass());
+ // moving IndVarSimplify here prevented removing the loop in perf_sumcartesian(10:-1:1)
+ PM->add(createLoopIdiomPass());
 #ifdef USE_POLLY
  // LCSSA (which has already run at this point due to the dependencies of the
  // above passes) introduces redundant phis that hinder Polly. Therefore we
@@ -196,46 +194,52 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level,
  PM->add(polly::createCodegenCleanupPass());
 #endif
  // LoopRotate strips metadata from terminator, so run LowerSIMD afterwards
- PM->add(createLowerSimdLoopPass());  // Annotate loop marked with "loopinfo" as LLVM parallel loop
- PM->add(createLICMPass()); // Hoist loop invariants
- PM->add(createLoopUnswitchPass()); // Unswitch loops.
+ PM->add(createLowerSimdLoopPass()); // Annotate loop marked with "loopinfo" as LLVM parallel loop
+ PM->add(createLICMPass());
+ PM->add(createLoopUnswitchPass());
  // Subsequent passes not stripping metadata from terminator
- PM->add(createInstructionCombiningPass());
- PM->add(createIndVarSimplifyPass()); // Canonicalize indvars
- PM->add(createLoopDeletionPass()); // Delete dead loops
- PM->add(createSimpleLoopUnrollPass()); // Unroll small loops
- //PM->add(createLoopStrengthReducePass()); // (jwb added)
+ PM->add(createInstSimplifyLegacyPass());
+ PM->add(createIndVarSimplifyPass());
+ PM->add(createLoopDeletionPass());
+ PM->add(createSimpleLoopUnrollPass());
 
  // Run our own SROA on heap objects before LLVM's
  PM->add(createAllocOptPass());
  // Re-run SROA after loop-unrolling (useful for small loops that operate,
  // over the structure of an aggregate)
- PM->add(createSROAPass()); // Break up aggregate allocas
- PM->add(createInstructionCombiningPass()); // Clean up after the unroller
- PM->add(createGVNPass()); // Remove redundancies
- PM->add(createMemCpyOptPass()); // Remove memcpy / form memset
- PM->add(createSCCPPass()); // Constant prop with SCCP
+ PM->add(createSROAPass());
+ // might not be necessary:
+ PM->add(createInstSimplifyLegacyPass());
+
+ PM->add(createGVNPass());
+ PM->add(createMemCpyOptPass());
+ PM->add(createSCCPPass());
 
  // Run instcombine after redundancy elimination to exploit opportunities
  // opened up by them.
+ // This needs to be InstCombine instead of InstSimplify to allow
+ // loops over Union-typed arrays to vectorize.
  PM->add(createInstructionCombiningPass());
- PM->add(createJumpThreadingPass()); // Thread jumps
- PM->add(createDeadStoreEliminationPass()); // Delete dead stores
+ PM->add(createJumpThreadingPass());
+ PM->add(createDeadStoreEliminationPass());
 
  // More dead allocation (store) deletion before loop optimization
+ // consider removing this:
  PM->add(createAllocOptPass());
  // see if all of the constant folding has exposed more loops
  // to simplification and deletion
  // this helps significantly with cleaning up iteration
- PM->add(createCFGSimplificationPass()); // Merge & remove BBs
- PM->add(createLoopIdiomPass());
- PM->add(createLoopDeletionPass()); // Delete dead loops
- PM->add(createJumpThreadingPass()); // Thread jumps
- PM->add(createSLPVectorizerPass()); // Vectorize straight-line code
- PM->add(createAggressiveDCEPass()); // Delete dead instructions
- PM->add(createInstructionCombiningPass()); // Clean up after SLP loop vectorizer
- PM->add(createLoopVectorizePass()); // Vectorize loops
- PM->add(createInstructionCombiningPass()); // Clean up after loop vectorizer
+ PM->add(createCFGSimplificationPass());
+ PM->add(createLoopDeletionPass());
+ PM->add(createInstructionCombiningPass());
+ PM->add(createLoopVectorizePass());
+ PM->add(createLoopLoadEliminationPass());
+ PM->add(createCFGSimplificationPass());
+ PM->add(createSLPVectorizerPass());
+ // might need this after LLVM 11:
+ //PM->add(createVectorCombinePass());
+
+ PM->add(createAggressiveDCEPass());
 
  if (lower_intrinsics) {
  // LowerPTLS removes an indirect call. As a result, it is likely to trigger
@@ -253,6 +257,7 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level,
  PM->add(createCFGSimplificationPass());
  }
  PM->add(createCombineMulAddPass());
+ PM->add(createDivRemPairsPass());
 }
 
 extern "C" JL_DLLEXPORT