Fix the InstCombine to reserve the VP metadata and sets correct call count.

[danielcdh]

Summary

Currently the VP metadata was dropped when InstCombine converts a call to direct call. This patch converts the VP metadata to branch_weights so that its hotness is recorded.

[davidxl]

lgtm

lib/Transforms/InstCombine/InstCombineCalls.cpp

4134 ↗

(On Diff #92968)

Add a comment here for the need of preserving vp data.

[danielcdh]

rebase and update

llvm/trunk/lib/IR/Metadata.cpp

assert((getOpcode() == Instruction::Br ||
          "Looking for branch weights on something besides branch");
 
   TotalVal = 0;
   auto *ProfileData = getMetadata(LLVMContext::MD_prof);
   if (!ProfileData)
     return false;
 
   auto *ProfDataName = dyn_cast<MDString>(ProfileData->getOperand(0));
-  if (!ProfDataName || !ProfDataName->getString().equals("branch_weights"))
+  if (!ProfDataName)
     return false;
 
+  if (ProfDataName->getString().equals("branch_weights")) {
     TotalVal = 0;
     for (unsigned i = 1; i < ProfileData->getNumOperands(); i++) {
       auto *V = mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(i));
       if (!V)
         return false;
       TotalVal += V->getValue().getZExtValue();
     }
     return true;
+  } else if (ProfDataName->getString().equals("VP") &&
+             ProfileData->getNumOperands() > 3) {
+    TotalVal = mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2))
+                   ->getValue()
+                   .getZExtValue();
+    return true;
+  }
+  return false;
 }
 
 void Instruction::clearMetadataHashEntries() {
   assert(hasMetadataHashEntry() && "Caller should check");
   getContext().pImpl->InstructionMetadata.erase(this);
   setHasMetadataHashEntry(false);
 }
 

llvm/trunk/lib/Transforms/InstCombine/InstCombineCalls.cpp

NC = Builder->CreateInvoke(Callee, II->getNormalDest(), II->getUnwindDest(),
                                Args, OpBundles);
     NC->takeName(II);
     cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
     cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
   } else {
     CallInst *CI = cast<CallInst>(Caller);
     NC = Builder->CreateCall(Callee, Args, OpBundles);
     NC->takeName(CI);
+    // Preserve the weight metadata for the new call instruction. The metadata
+    // is used by SamplePGO to check callsite's hotness.
+    uint64_t W;
+    if (CI->extractProfTotalWeight(W))
+      NC->setProfWeight(W);
+
     cast<CallInst>(NC)->setTailCallKind(CI->getTailCallKind());
     cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
     cast<CallInst>(NC)->setAttributes(NewCallerPAL);
   }
 
   // Insert a cast of the return type as necessary.
   Value *NV = NC;
   if (OldRetTy != NV->getType() && !Caller->use_empty()) {

llvm/trunk/test/Transforms/InstCombine/cast-call-combine-prof.ll

+; RUN: opt -instcombine -inline -S -inline-threshold=0 -hot-callsite-threshold=100 < %s | FileCheck %s
+; Checks if VP profile is used for hotness checks in inlining after instcombine
+; converted the call to a direct call.
+
+declare void @bar(i16 *)
+
+define void @foo(i16* %a) {
+  call void @bar(i16* %a)
+  call void @bar(i16* %a)
+  ret void
+}
+
+; CHECK-LABEL: @test()
+; CHECK-NEXT: call void @bar
+; CHECK-NEXT: call void @bar
+define void @test() {
+  call void bitcast (void (i16*)* @foo to void (i8*)*) (i8* null), !prof !0
+  ret void
+}
+
+!0 = !{!"VP", i32 0, i64 2000, i64 -3913987384944532146, i64 2000}
+
+!llvm.module.flags = !{!1}
+
+!1 = !{i32 1, !"ProfileSummary", !2}
+!2 = !{!3, !4, !5, !6, !7, !8, !9, !10}
+!3 = !{!"ProfileFormat", !"InstrProf"}
+!4 = !{!"TotalCount", i64 10000}
+!5 = !{!"MaxCount", i64 1000}
+!6 = !{!"MaxInternalCount", i64 1}
+!7 = !{!"MaxFunctionCount", i64 1000}
+!8 = !{!"NumCounts", i64 3}
+!9 = !{!"NumFunctions", i64 3}
+!10 = !{!"DetailedSummary", !11}
+!11 = !{!12, !13, !14}
+!12 = !{i32 10000, i64 1000, i32 1}
+!13 = !{i32 999000, i64 1000, i32 1}
+!14 = !{i32 999999, i64 1, i32 2}