[PGO][PGSO] Add profile guided size optimization to loop vectorization legality.

[davidxl]

llvm/lib/Transforms/Vectorize/LoopVectorizationLegality.cpp

418

should the check of F->hasOptSize() also folded into llvm::shouldOptimizeForSize()?

[hjyamauchi]

llvm/lib/Transforms/Vectorize/LoopVectorizationLegality.cpp

418

That's actually been suggested before and seems to make sense, and is in my todo list for a while. But it'd be a different patch because it'd be an API change and apply to many sites that have been instrumented with llvm::shouldOptimizeForSize() so far.

[davidxl]

lgtm

[hjyamauchi]

Rebase.

[aeubanks]

I've bisected a Chromium ThinLTO failing assert to this change, could you take a look?
https://crbug.com/1106813

[aeubanks]

Reverted in 0dfa4a83fad43f95c90dd67cb23f63baaa907bd1.

[hjyamauchi]

Reproduced https://crbug.com/1106813.

The problem was that we were querying the profile of the newly created loop, as opposed to the original loop in InnerLoopVectorizer.

Querying against the original loop (and saving it in the constructor before the transformations because the profile of the original loop could change as it is used as the fallback loop) should fix the issue.

[hjyamauchi]

Update after a fix for https://crbug.com/1106813

[hjyamauchi]

PTAL. This is the diff of the patches before/after the latest fix.

70c70,76
< +        BFI(BFI), PSI(PSI) {}
---
> +        BFI(BFI), PSI(PSI) {
> +    // Query this against the original loop and save it here because the profile
> +    // of the original loop header may change as the transformation happens.
> +    OptForSizeBasedOnProfile = llvm::shouldOptimizeForSize(
> +        OrigLoop->getHeader(), PSI, BFI, PGSOQueryType::IRPass);
> +  }
> +
74c80
< @@ -779,6 +781,10 @@ protected:
---
> @@ -779,6 +787,14 @@ protected:
81a88,91
> +
> +  // Whether this loop should be optimized for size based on profile guided size
> +  // optimizatios.
> +  bool OptForSizeBasedOnProfile;
85c95
< @@ -789,9 +795,10 @@ public:
---
> @@ -789,9 +805,10 @@ public:
98c108
< @@ -2754,7 +2761,9 @@ void InnerLoopVectorizer::emitSCEVChecks(Loop *L, BasicBlock *Bypass) {
---
> @@ -2754,7 +2771,8 @@ void InnerLoopVectorizer::emitSCEVChecks(Loop *L, BasicBlock *Bypass) {
104,105c114
< +           llvm::shouldOptimizeForSize(L->getHeader(), PSI, BFI,
< +                                       PGSOQueryType::IRPass)) &&
---
> +           OptForSizeBasedOnProfile) &&
109c118
< @@ -2800,7 +2809,9 @@ void InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass) {
---
> @@ -2800,7 +2818,7 @@ void InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass) {
114,116c123
< +  if (MemCheckBlock->getParent()->hasOptSize() ||
< +      llvm::shouldOptimizeForSize(L->getHeader(), PSI, BFI,
< +                                  PGSOQueryType::IRPass)) {
---
> +  if (MemCheckBlock->getParent()->hasOptSize() || OptForSizeBasedOnProfile) {

[davidxl]

thanks for the fix. looks good.

llvm/include/llvm/Transforms/Vectorize/LoopVectorizationLegality.h

 class LoopVectorizationLegality {
 public:
   LoopVectorizationLegality(
       Loop *L, PredicatedScalarEvolution &PSE, DominatorTree *DT,
       TargetTransformInfo *TTI, TargetLibraryInfo *TLI, AAResults *AA,
       Function *F, std::function<const LoopAccessInfo &(Loop &)> *GetLAA,
       LoopInfo *LI, OptimizationRemarkEmitter *ORE,
       LoopVectorizationRequirements *R, LoopVectorizeHints *H, DemandedBits *DB,
-      AssumptionCache *AC)
+      AssumptionCache *AC, BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI)
       : TheLoop(L), LI(LI), PSE(PSE), TTI(TTI), TLI(TLI), DT(DT),
-        GetLAA(GetLAA), ORE(ORE), Requirements(R), Hints(H), DB(DB), AC(AC) {}
+        GetLAA(GetLAA), ORE(ORE), Requirements(R), Hints(H), DB(DB), AC(AC),
+        BFI(BFI), PSI(PSI) {}
 
   /// ReductionList contains the reduction descriptors for all
   /// of the reductions that were found in the loop.
   using ReductionList = MapVector<PHINode *, RecurrenceDescriptor>;
 
   /// InductionList saves induction variables and maps them to the
   /// induction descriptor.
   using InductionList = MapVector<PHINode *, InductionDescriptor>;
private:
 
   /// While vectorizing these instructions we have to generate a
   /// call to the appropriate masked intrinsic
   SmallPtrSet<const Instruction *, 8> MaskedOp;
 
   /// Assume instructions in predicated blocks must be dropped if the CFG gets
   /// flattened.
   SmallPtrSet<Instruction *, 8> ConditionalAssumes;
+
+  /// BFI and PSI are used to check for profile guided size optimizations.
+  BlockFrequencyInfo *BFI;
+  ProfileSummaryInfo *PSI;
 };
 
 } // namespace llvm
 
 #endif // LLVM_TRANSFORMS_VECTORIZE_LOOPVECTORIZATIONLEGALITY_H

llvm/lib/Transforms/Vectorize/LoopVectorizationLegality.cpp

 //
 #include "llvm/Transforms/Vectorize/LoopVectorizationLegality.h"
 #include "llvm/Analysis/Loads.h"
 #include "llvm/Analysis/LoopInfo.h"
 #include "llvm/Analysis/ValueTracking.h"
 #include "llvm/Analysis/VectorUtils.h"
 #include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/PatternMatch.h"
+#include "llvm/Transforms/Utils/SizeOpts.h"
 #include "llvm/Transforms/Vectorize/LoopVectorize.h"
 
 using namespace llvm;
 using namespace PatternMatch;
 
 #define LV_NAME "loop-vectorize"
 #define DEBUG_TYPE LV_NAME
 
if (!AllowedExit.count(Inst))
     }
   return false;
 }
 
 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
   const ValueToValueMap &Strides =
       getSymbolicStrides() ? *getSymbolicStrides() : ValueToValueMap();
 
-  bool CanAddPredicate = !TheLoop->getHeader()->getParent()->hasOptSize();
+  Function *F = TheLoop->getHeader()->getParent();
+  bool OptForSize = F->hasOptSize() ||
+                    llvm::shouldOptimizeForSize(TheLoop->getHeader(), PSI, BFI,
+                                                PGSOQueryType::IRPass);
+  bool CanAddPredicate = !OptForSize;
   int Stride = getPtrStride(PSE, Ptr, TheLoop, Strides, CanAddPredicate, false);
   if (Stride == 1 || Stride == -1)
     return Stride;
   return 0;
 }
 
 bool LoopVectorizationLegality::isUniform(Value *V) {
   return LAI->isUniform(V);

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

 class InnerLoopVectorizer {
 public:
   InnerLoopVectorizer(Loop *OrigLoop, PredicatedScalarEvolution &PSE,
                       LoopInfo *LI, DominatorTree *DT,
                       const TargetLibraryInfo *TLI,
                       const TargetTransformInfo *TTI, AssumptionCache *AC,
                       OptimizationRemarkEmitter *ORE, unsigned VecWidth,
                       unsigned UnrollFactor, LoopVectorizationLegality *LVL,
-                      LoopVectorizationCostModel *CM)
+                      LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI,
+                      ProfileSummaryInfo *PSI)
       : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
         AC(AC), ORE(ORE), VF(VecWidth), UF(UnrollFactor),
         Builder(PSE.getSE()->getContext()),
-        VectorLoopValueMap(UnrollFactor, VecWidth), Legal(LVL), Cost(CM) {}
+        VectorLoopValueMap(UnrollFactor, VecWidth), Legal(LVL), Cost(CM),
+        BFI(BFI), PSI(PSI) {}
   virtual ~InnerLoopVectorizer() = default;
 
   /// Create a new empty loop. Unlink the old loop and connect the new one.
   /// Return the pre-header block of the new loop.
   BasicBlock *createVectorizedLoopSkeleton();
 
   /// Widen a single instruction within the innermost loop.
   void widenInstruction(Instruction &I, VPUser &Operands,
protected:
 
   // Holds the end values for each induction variable. We save the end values
   // so we can later fix-up the external users of the induction variables.
   DenseMap<PHINode *, Value *> IVEndValues;
 
   // Vector of original scalar PHIs whose corresponding widened PHIs need to be
   // fixed up at the end of vector code generation.
   SmallVector<PHINode *, 8> OrigPHIsToFix;
+
+  /// BFI and PSI are used to check for profile guided size optimizations.
+  BlockFrequencyInfo *BFI;
+  ProfileSummaryInfo *PSI;
 };
 
 class InnerLoopUnroller : public InnerLoopVectorizer {
 public:
   InnerLoopUnroller(Loop *OrigLoop, PredicatedScalarEvolution &PSE,
                     LoopInfo *LI, DominatorTree *DT,
                     const TargetLibraryInfo *TLI,
                     const TargetTransformInfo *TTI, AssumptionCache *AC,
                     OptimizationRemarkEmitter *ORE, unsigned UnrollFactor,
                     LoopVectorizationLegality *LVL,
-                    LoopVectorizationCostModel *CM)
+                    LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI,
+                    ProfileSummaryInfo *PSI)
       : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, 1,
-                            UnrollFactor, LVL, CM) {}
+                            UnrollFactor, LVL, CM, BFI, PSI) {}
 
 private:
   Value *getBroadcastInstrs(Value *V) override;
   Value *getStepVector(Value *Val, int StartIdx, Value *Step,
                        Instruction::BinaryOps Opcode =
                        Instruction::BinaryOpsEnd) override;
   Value *reverseVector(Value *Vec) override;
 };
SCEVExpander Exp(*PSE.getSE(), Bypass->getModule()->getDataLayout(),
                    "scev.check");
   Value *SCEVCheck = Exp.expandCodeForPredicate(
       &PSE.getUnionPredicate(), SCEVCheckBlock->getTerminator());
 
   if (auto *C = dyn_cast<ConstantInt>(SCEVCheck))
     if (C->isZero())
       return;
 
-  assert(!SCEVCheckBlock->getParent()->hasOptSize() &&
+  assert(!(SCEVCheckBlock->getParent()->hasOptSize() ||
+           llvm::shouldOptimizeForSize(L->getHeader(), PSI, BFI,
+                                       PGSOQueryType::IRPass)) &&
          "Cannot SCEV check stride or overflow when optimizing for size");
 
   SCEVCheckBlock->setName("vector.scevcheck");
   // Create new preheader for vector loop.
   LoopVectorPreHeader =
       SplitBlock(SCEVCheckBlock, SCEVCheckBlock->getTerminator(), DT, LI,
                  nullptr, "vector.ph");
 
void InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass) {
   Instruction *FirstCheckInst;
   Instruction *MemRuntimeCheck;
   std::tie(FirstCheckInst, MemRuntimeCheck) =
       addRuntimeChecks(MemCheckBlock->getTerminator(), OrigLoop,
                        RtPtrChecking.getChecks(), RtPtrChecking.getSE());
   assert(MemRuntimeCheck && "no RT checks generated although RtPtrChecking "
                             "claimed checks are required");
 
-  if (MemCheckBlock->getParent()->hasOptSize()) {
+  if (MemCheckBlock->getParent()->hasOptSize() ||
+      llvm::shouldOptimizeForSize(L->getHeader(), PSI, BFI,
+                                  PGSOQueryType::IRPass)) {
     assert(Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled &&
            "Cannot emit memory checks when optimizing for size, unless forced "
            "to vectorize.");
     ORE->emit([&]() {
       return OptimizationRemarkAnalysis(DEBUG_TYPE, "VectorizationCodeSize",
                                         L->getStartLoc(), L->getHeader())
              << "Code-size may be reduced by not forcing "
                 "vectorization, or by source-code modifications "
static bool processLoopInVPlanNativePath(
   // Also, do not attempt to vectorize if no vector code will be produced.
   if (VPlanBuildStressTest || EnableVPlanPredication ||
       VectorizationFactor::Disabled() == VF)
     return false;
 
   LVP.setBestPlan(VF.Width, 1);
 
   InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, 1, LVL,
-                         &CM);
+                         &CM, BFI, PSI);
   LLVM_DEBUG(dbgs() << "Vectorizing outer loop in \""
                     << L->getHeader()->getParent()->getName() << "\"\n");
   LVP.executePlan(LB, DT);
 
   // Mark the loop as already vectorized to avoid vectorizing again.
   Hints.setAlreadyVectorized();
 
   assert(!verifyFunction(*L->getHeader()->getParent(), &dbgs()));
if (!Hints.allowVectorization(F, L, VectorizeOnlyWhenForced)) {
     return false;
   }
 
   PredicatedScalarEvolution PSE(*SE, *L);
 
   // Check if it is legal to vectorize the loop.
   LoopVectorizationRequirements Requirements(*ORE);
   LoopVectorizationLegality LVL(L, PSE, DT, TTI, TLI, AA, F, GetLAA, LI, ORE,
-                                &Requirements, &Hints, DB, AC);
+                                &Requirements, &Hints, DB, AC, BFI, PSI);
   if (!LVL.canVectorize(EnableVPlanNativePath)) {
     LLVM_DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
     Hints.emitRemarkWithHints();
     return false;
   }
 
   // Check the function attributes and profiles to find out if this function
   // should be optimized for size.
#endif /* NDEBUG */
   using namespace ore;
   bool DisableRuntimeUnroll = false;
   MDNode *OrigLoopID = L->getLoopID();
 
   if (!VectorizeLoop) {
     assert(IC > 1 && "interleave count should not be 1 or 0");
     // If we decided that it is not legal to vectorize the loop, then
     // interleave it.
-    InnerLoopUnroller Unroller(L, PSE, LI, DT, TLI, TTI, AC, ORE, IC, &LVL,
-                               &CM);
+    InnerLoopUnroller Unroller(L, PSE, LI, DT, TLI, TTI, AC, ORE, IC, &LVL, &CM,
+                               BFI, PSI);
     LVP.executePlan(Unroller, DT);
 
     ORE->emit([&]() {
       return OptimizationRemark(LV_NAME, "Interleaved", L->getStartLoc(),
                                 L->getHeader())
              << "interleaved loop (interleaved count: "
              << NV("InterleaveCount", IC) << ")";
     });
   } else {
     // If we decided that it is *legal* to vectorize the loop, then do it.
     InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, IC,
-                           &LVL, &CM);
+                           &LVL, &CM, BFI, PSI);
     LVP.executePlan(LB, DT);
     ++LoopsVectorized;
 
     // Add metadata to disable runtime unrolling a scalar loop when there are
     // no runtime checks about strides and memory. A scalar loop that is
     // rarely used is not worth unrolling.
     if (!LB.areSafetyChecksAdded())
       DisableRuntimeUnroll = true;

llvm/test/Transforms/LoopVectorize/optsize.ll

for.body29:
   %idxprom34 = zext i16 %add33 to i32
   %arrayidx35 = getelementptr [2592 x i16], [2592 x i16] * @cm_array, i32 0, i32 %idxprom34
   store i16 0, i16 * %arrayidx35, align 1
   %inc37 = add i16 %i24.0170, 1
   %cmp26 = icmp ult i16 %inc37, 756
   br i1 %cmp26, label %for.body29, label %for.cond.cleanup28
 }
 
+define void @pr43371_pgso() !prof !14 {
+;
+; CHECK-LABEL: @pr43371_pgso
+; CHECK-NOT:   vector.scevcheck
+;
+; We do not want to generate SCEV predicates when optimising for size, because
+; that will lead to extra code generation such as the SCEV overflow runtime
+; checks. Not generating SCEV predicates can still result in vectorisation as
+; the non-consecutive loads/stores can be scalarized:
+;
+; CHECK: vector.body:
+; CHECK: store i16 0, i16* %{{.*}}, align 1
+; CHECK: store i16 0, i16* %{{.*}}, align 1
+; CHECK: br i1 {{.*}}, label %vector.body
+;
+entry:
+  br label %for.body29
+
+for.cond.cleanup28:
+  unreachable
+
+for.body29:
+  %i24.0170 = phi i16 [ 0, %entry], [ %inc37, %for.body29]
+  %add33 = add i16 undef, %i24.0170
+  %idxprom34 = zext i16 %add33 to i32
+  %arrayidx35 = getelementptr [2592 x i16], [2592 x i16] * @cm_array, i32 0, i32 %idxprom34
+  store i16 0, i16 * %arrayidx35, align 1
+  %inc37 = add i16 %i24.0170, 1
+  %cmp26 = icmp ult i16 %inc37, 756
+  br i1 %cmp26, label %for.body29, label %for.cond.cleanup28
+}
+
 ; PR45526: don't vectorize with fold-tail if first-order-recurrence is live-out.
 ;
 define i32 @pr45526() optsize {
 ;
 ; CHECK-LABEL: @pr45526
 ; CHECK-NEXT: entry:
 ; CHECK-NEXT:   br label %loop
 ; CHECK-EMPTY:
loop:
   %pivPlus1 = add nuw nsw i32 %piv, 1
   %cond = icmp ult i32 %piv, 510
   br i1 %cond, label %loop, label %exit
 
 exit:
   ret i32 %for
 }
 
+define i32 @pr45526_pgso() !prof !14 {
+;
+; CHECK-LABEL: @pr45526_pgso
+; CHECK-NEXT: entry:
+; CHECK-NEXT:   br label %loop
+; CHECK-EMPTY:
+; CHECK-NEXT: loop:
+; CHECK-NEXT:   %piv = phi i32 [ 0, %entry ], [ %pivPlus1, %loop ]
+; CHECK-NEXT:   %for = phi i32 [ 5, %entry ], [ %pivPlus1, %loop ]
+; CHECK-NEXT:   %pivPlus1 = add nuw nsw i32 %piv, 1
+; CHECK-NEXT:   %cond = icmp ult i32 %piv, 510
+; CHECK-NEXT:   br i1 %cond, label %loop, label %exit
+; CHECK-EMPTY:
+; CHECK-NEXT: exit:
+; CHECK-NEXT:   %for.lcssa = phi i32 [ %for, %loop ]
+; CHECK-NEXT:   ret i32 %for.lcssa
+;
+entry:
+  br label %loop
+
+loop:
+  %piv = phi i32 [ 0, %entry ], [ %pivPlus1, %loop ]
+  %for = phi i32 [ 5, %entry ], [ %pivPlus1, %loop ]
+  %pivPlus1 = add nuw nsw i32 %piv, 1
+  %cond = icmp ult i32 %piv, 510
+  br i1 %cond, label %loop, label %exit
+
+exit:
+  ret i32 %for
+}
+
 ; PR46228: Vectorize w/o versioning for unit stride under optsize and enabled
 ; vectorization.
 
 ; NOTE: Some assertions have been autogenerated by utils/update_test_checks.py
 define void @stride1(i16* noalias %B, i32 %BStride) optsize {
 ; CHECK-LABEL: @stride1(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]

 ; CHECK-NEXT:    [[TMP6:%.*]] = extractelement <2 x i32> [[TMP0]], i32 1
 ; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr inbounds i16, i16* [[B]], i32 [[TMP6]]
 ; CHECK-NEXT:    store i16 42, i16* [[TMP7]], align 4
 ; CHECK-NEXT:    br label [[PRED_STORE_CONTINUE2]]
 ; CHECK:       pred.store.continue2:
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i32 [[INDEX]], 2
 ; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <2 x i32> [[VEC_IND]], <i32 2, i32 2>
 ; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i32 [[INDEX_NEXT]], 1026
-; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !19
+; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !21
 ; CHECK:       middle.block:
 ; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK:       for.end:
 ; CHECK-NEXT:    ret void
 ;
 ; PGSO-LABEL: @stride1(
 ; PGSO-NEXT:  entry: