[NewPassManager] Add tuning option: ForgetAllSCEVInLoopUnroll [NFC].

[asbirlea]

Summary

Mirror tuning option from old pass manager in new pass manager.

[asbirlea]

Update comment to include flag.

[chandlerc]

This seems... *really* unlikely to be an important tuning flag for library clients. It seems likely added to allow some limited experimentation. Maybe we can just not add it to the new PM?

[asbirlea]

It is currently turned on by default for Halide, which is still using the old pass manager. It makes a big difference for them in compile times, so switching them to the new pass manager will cause regressions without this flag (see https://github.com/halide/Halide/blob/master/src/CodeGen_LLVM.cpp#L1218).

Alternatively, we could enable it by default, since it benefits clang as well (see timings in D60144).

I don't have a preference either way.

[chandlerc]

Ok, I'm fine with this going in temporarily while we try to flip this flag acros sthe board (or figure out a heuristic that lets us get the best of both worlds). So, LGTM, but yeah, let's try to remove this too.

include/llvm/Passes/PassBuilder.h

public:
   /// Tuning option to set loop interleaving on/off. Its default value is that
   /// of the flag: `-interleave-loops`.
   bool LoopInterleaving;
 
   /// Tuning option to enable/disable loop vectorization. Its default value is
   /// that of the flag: `-vectorize-loops`.
   bool LoopVectorization;
 
+  /// Tuning option to forget all SCEV loops in LoopUnroll. Its default value
+  /// is that of the flag: `-forget-scev-loop-unroll`.
+  bool ForgetAllSCEVInLoopUnroll;
+
   /// Tuning option to cap the number of calls to retrive clobbering accesses in
   /// MemorySSA, in LICM.
   unsigned LicmMssaOptCap;
 
   /// Tuning option to disable promotion to scalars in LICM with MemorySSA, if
   /// the number of access is too large.
   unsigned LicmMssaNoAccForPromotionCap;
 };

include/llvm/Transforms/Scalar/LoopUnrollPass.h

 #define LLVM_TRANSFORMS_SCALAR_LOOPUNROLLPASS_H
 
 #include "llvm/ADT/Optional.h"
 #include "llvm/Analysis/LoopAnalysisManager.h"
 #include "llvm/IR/PassManager.h"
 
 namespace llvm {
 
+extern cl::opt<bool> ForgetSCEVInLoopUnroll;
+
 class Function;
 class Loop;
 class LPMUpdater;
 
 /// Loop unroll pass that only does full loop unrolling.
 class LoopFullUnrollPass : public PassInfoMixin<LoopFullUnrollPass> {
   const int OptLevel;
 
   /// If false, use a cost model to determine whether unrolling of a loop is
   /// profitable. If true, only loops that explicitly request unrolling via
   /// metadata are considered. All other loops are skipped.
   const bool OnlyWhenForced;
 
+  /// If true, forget all loops when unrolling. If false, forget top-most loop
+  /// of the currently processed loops, which removes one entry at a time from
+  /// the internal SCEV records. For large loops, the former is faster.
+  const bool ForgetSCEV;
+
 public:
-  explicit LoopFullUnrollPass(int OptLevel = 2, bool OnlyWhenForced = false)
-      : OptLevel(OptLevel), OnlyWhenForced(OnlyWhenForced) {}
+  explicit LoopFullUnrollPass(int OptLevel = 2, bool OnlyWhenForced = false,
+                              bool ForgetSCEV = false)
+      : OptLevel(OptLevel), OnlyWhenForced(OnlyWhenForced),
+        ForgetSCEV(ForgetSCEV) {}
 
   PreservedAnalyses run(Loop &L, LoopAnalysisManager &AM,
                         LoopStandardAnalysisResults &AR, LPMUpdater &U);
 };
 
 /// A set of parameters used to control various transforms performed by the
 /// LoopUnroll pass. Each of the boolean parameters can be set to:
 ///      true - enabling the transformation.
struct LoopUnrollOptions {
   Optional<bool> AllowUpperBound;
   int OptLevel;
 
   /// If false, use a cost model to determine whether unrolling of a loop is
   /// profitable. If true, only loops that explicitly request unrolling via
   /// metadata are considered. All other loops are skipped.
   bool OnlyWhenForced;
 
-  LoopUnrollOptions(int OptLevel = 2, bool OnlyWhenForced = false)
-      : OptLevel(OptLevel), OnlyWhenForced(OnlyWhenForced) {}
+  /// If true, forget all loops when unrolling. If false, forget top-most loop
+  /// of the currently processed loops, which removes one entry at a time from
+  /// the internal SCEV records. For large loops, the former is faster.
+  const bool ForgetSCEV;
+
+  LoopUnrollOptions(int OptLevel = 2, bool OnlyWhenForced = false,
+                    bool ForgetSCEV = false)
+      : OptLevel(OptLevel), OnlyWhenForced(OnlyWhenForced),
+        ForgetSCEV(ForgetSCEV) {}
 
   /// Enables or disables partial unrolling. When disabled only full unrolling
   /// is allowed.
   LoopUnrollOptions &setPartial(bool Partial) {
     AllowPartial = Partial;
     return *this;
   }
 

lib/Passes/PassBuilder.cpp

 // profile loading.
 static cl::opt<bool>
     EnableCHR("enable-chr-npm", cl::init(true), cl::Hidden,
               cl::desc("Enable control height reduction optimization (CHR)"));
 
 PipelineTuningOptions::PipelineTuningOptions() {
   LoopInterleaving = EnableLoopInterleaving;
   LoopVectorization = EnableLoopVectorization;
+  ForgetAllSCEVInLoopUnroll = ForgetSCEVInLoopUnroll;
   LicmMssaOptCap = SetLicmMssaOptCap;
   LicmMssaNoAccForPromotionCap = SetLicmMssaNoAccForPromotionCap;
 }
 
 extern cl::opt<bool> EnableHotColdSplit;
 extern cl::opt<bool> EnableOrderFileInstrumentation;
 
 extern cl::opt<bool> FlattenedProfileUsed;
for (auto &C : LateLoopOptimizationsEPCallbacks)
     C(LPM2, Level);
 
   LPM2.addPass(LoopDeletionPass());
   // Do not enable unrolling in PreLinkThinLTO phase during sample PGO
   // because it changes IR to makes profile annotation in back compile
   // inaccurate.
   if (Phase != ThinLTOPhase::PreLink || !PGOOpt ||
       PGOOpt->Action != PGOOptions::SampleUse)
-    LPM2.addPass(LoopFullUnrollPass(Level));
+    LPM2.addPass(
+        LoopFullUnrollPass(Level, false, PTO.ForgetAllSCEVInLoopUnroll));
 
   for (auto &C : LoopOptimizerEndEPCallbacks)
     C(LPM2, Level);
 
   // We provide the opt remark emitter pass for LICM to use. We only need to do
   // this once as it is immutable.
   FPM.addPass(RequireAnalysisPass<OptimizationRemarkEmitterAnalysis, Function>());
   FPM.addPass(createFunctionToLoopPassAdaptor(std::move(LPM1), DebugLogging));
ModulePassManager PassBuilder::buildModuleOptimizationPipeline(
   // FIXME: It would be really good to use a loop-integrated instruction
   // combiner for cleanup here so that the unrolling and LICM can be pipelined
   // across the loop nests.
   // We do UnrollAndJam in a separate LPM to ensure it happens before unroll
   if (EnableUnrollAndJam) {
     OptimizePM.addPass(
         createFunctionToLoopPassAdaptor(LoopUnrollAndJamPass(Level)));
   }
-  OptimizePM.addPass(LoopUnrollPass(LoopUnrollOptions(Level)));
+  OptimizePM.addPass(LoopUnrollPass(
+      LoopUnrollOptions(Level, false, PTO.ForgetAllSCEVInLoopUnroll)));
   OptimizePM.addPass(WarnMissedTransformationsPass());
   OptimizePM.addPass(InstCombinePass());
   OptimizePM.addPass(RequireAnalysisPass<OptimizationRemarkEmitterAnalysis, Function>());
   OptimizePM.addPass(createFunctionToLoopPassAdaptor(
       LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap),
       DebugLogging));
 
   // Now that we've vectorized and unrolled loops, we may have more refined

lib/Transforms/IPO/PassManagerBuilder.cpp

 #include "llvm/Transforms/IPO/FunctionAttrs.h"
 #include "llvm/Transforms/IPO/InferFunctionAttrs.h"
 #include "llvm/Transforms/InstCombine/InstCombine.h"
 #include "llvm/Transforms/Instrumentation.h"
 #include "llvm/Transforms/Scalar.h"
 #include "llvm/Transforms/Scalar/GVN.h"
 #include "llvm/Transforms/Scalar/InstSimplifyPass.h"
 #include "llvm/Transforms/Scalar/LICM.h"
+#include "llvm/Transforms/Scalar/LoopUnrollPass.h"
 #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h"
 #include "llvm/Transforms/Utils.h"
 #include "llvm/Transforms/Vectorize.h"
 #include "llvm/Transforms/Vectorize/LoopVectorize.h"
 
 using namespace llvm;
 
 static cl::opt<bool>
cl::opt<bool> FlattenedProfileUsed(
     "flattened-profile-used", cl::init(false), cl::Hidden,
     cl::desc("Indicate the sample profile being used is flattened, i.e., "
              "no inline hierachy exists in the profile. "));
 
 cl::opt<bool> EnableOrderFileInstrumentation(
     "enable-order-file-instrumentation", cl::init(false), cl::Hidden,
     cl::desc("Enable order file instrumentation (default = off)"));
 
-cl::opt<bool> ForgetSCEVInLoopUnroll(
-    "forget-scev-loop-unroll", cl::init(false), cl::Hidden,
-    cl::desc("Forget everything in SCEV when doing LoopUnroll, instead of just"
-             " the current top-most loop. This is somtimes preferred to reduce"
-             " compile time."));
-
 PassManagerBuilder::PassManagerBuilder() {
     OptLevel = 2;
     SizeLevel = 0;
     LibraryInfo = nullptr;
     Inliner = nullptr;
     DisableUnrollLoops = false;
     SLPVectorize = RunSLPVectorization;
     LoopVectorize = EnableLoopVectorization;

lib/Transforms/Scalar/LoopUnrollPass.cpp

 #include <string>
 #include <tuple>
 #include <utility>
 
 using namespace llvm;
 
 #define DEBUG_TYPE "loop-unroll"
 
+cl::opt<bool> llvm::ForgetSCEVInLoopUnroll(
+    "forget-scev-loop-unroll", cl::init(false), cl::Hidden,
+    cl::desc("Forget everything in SCEV when doing LoopUnroll, instead of just"
+             " the current top-most loop. This is somtimes preferred to reduce"
+             " compile time."));
+
 static cl::opt<unsigned>
     UnrollThreshold("unroll-threshold", cl::Hidden,
                     cl::desc("The cost threshold for loop unrolling"));
 
 static cl::opt<unsigned> UnrollPartialThreshold(
     "unroll-partial-threshold", cl::Hidden,
     cl::desc("The cost threshold for partial loop unrolling"));
 
else
     OldLoops.insert(AR.LI.begin(), AR.LI.end());
 
   std::string LoopName = L.getName();
 
   bool Changed =
       tryToUnrollLoop(&L, AR.DT, &AR.LI, AR.SE, AR.TTI, AR.AC, *ORE,
                       /*BFI*/ nullptr, /*PSI*/ nullptr,
                       /*PreserveLCSSA*/ true, OptLevel, OnlyWhenForced,
-                      /*ForgetAllSCEV*/ false, /*Count*/ None,
+                      ForgetSCEV, /*Count*/ None,
                       /*Threshold*/ None, /*AllowPartial*/ false,
                       /*Runtime*/ false, /*UpperBound*/ false,
                       /*AllowPeeling*/ false) != LoopUnrollResult::Unmodified;
   if (!Changed)
     return PreservedAnalyses::all();
 
   // The parent must not be damaged by unrolling!
 #ifndef NDEBUG
#endif
     if (PSI && PSI->hasHugeWorkingSetSize())
       LocalAllowPeeling = false;
     std::string LoopName = L.getName();
     // The API here is quite complex to call and we allow to select some
     // flavors of unrolling during construction time (by setting UnrollOpts).
     LoopUnrollResult Result = tryToUnrollLoop(
         &L, DT, &LI, SE, TTI, AC, ORE, BFI, PSI,
         /*PreserveLCSSA*/ true, UnrollOpts.OptLevel, UnrollOpts.OnlyWhenForced,
-        /*ForgetAllSCEV*/ false, /*Count*/ None,
+        UnrollOpts.ForgetSCEV, /*Count*/ None,
         /*Threshold*/ None, UnrollOpts.AllowPartial, UnrollOpts.AllowRuntime,
         UnrollOpts.AllowUpperBound, LocalAllowPeeling);
     Changed |= Result != LoopUnrollResult::Unmodified;
 
     // The parent must not be damaged by unrolling!
 #ifndef NDEBUG
     if (Result != LoopUnrollResult::Unmodified && ParentL)
       ParentL->verifyLoop();