[PSI] Use zero function entry count info when getting profile count for calls

[scui]

Summary

Currently the BFI basic block ProfileCount is used when trying to getProfileCount for callsites. There are situations this count information may be inaccurate, for example, when the callsite code region is a result of inlining of a function without profile information to the caller which is profiled. This is to use the function entry count information from the called function when it contains the profile information and its function entry count is zero.

[davidxl]

I assume this is with instrumentation based FDO. Why are there functions without profile?

llvm/lib/Analysis/ProfileSummaryInfo.cpp

93

Put this in the else case for the following if(BFI) {} check below.

[scui]

In D158664#4611498, @davidxl wrote:

I assume this is with instrumentation based FDO. Why are there functions without profile?

Yes. this is with instrumentation based FDO. This can happen when customers are not profiling all the files of their applications.

llvm/lib/Analysis/ProfileSummaryInfo.cpp

if (hasSampleProfile()) {
     // instruction. If it is present, determine hotness solely based on that,
     // since the sampled entry count may not be accurate. If there is no
     // annotated on the instruction, return std::nullopt.
     uint64_t TotalCount;
     if (Call.extractProfTotalWeight(TotalCount))
       return TotalCount;
     return std::nullopt;
   }
+  if (auto *Called = Call.getCalledFunction()) {
+    auto CalledCount = Called->getEntryCount();
+    if (CalledCount && !CalledCount->getCount())
+      return 0;
+  }
   if (BFI)
     return BFI->getBlockProfileCount(Call.getParent(), AllowSynthetic);
   return std::nullopt;
 }
 
 bool ProfileSummaryInfo::isFunctionHotnessUnknown(const Function &F) const {
   assert(hasPartialSampleProfile() && "Expect partial sample profile");
   return !F.getEntryCount();

llvm/test/Transforms/Inline/inline-zero-count-call.ll

+; RUN: opt < %s -passes='require<profile-summary>,cgscc(inline)' -S | FileCheck %s
+
+; This tests a call to a function with zero entry count while no profile data
+; is available for the basic block of the call site.
+
+define void  @noprofile_f(i32 %n) "function-inline-cost"="15" {
+; CHECK-LABEL: define void @noprofile_f
+; CHECK:       cond_false:
+; CHECK-NEXT:    call void @zero_count_f()
+; CHECK-NEXT:    ret void
+;
+  %cond = icmp sle i32 %n, 10
+  br i1 %cond, label %cond_true, label %cond_false
+cond_true:
+  call void @extern_f()
+  ret void
+cond_false:
+  call void @zero_count_f()
+  ret void
+}
+
+define void @pprofile_f(i32 %n) !prof !16 {
+; CHECK-LABEL: define void @pprofile_f
+; CHECK:       cond_false:
+; CHECK-NEXT:    call void @zero_count_f()
+; CHECK-NEXT:    ret void
+;
+  %cond = icmp sle i32 %n, 10
+  br i1 %cond, label %cond_true, label %cond_false
+cond_true:
+  call void @extern_f()
+  ret void
+cond_false:
+  call void @zero_count_f()
+  ret void
+}
+
+define void @zero_count_f() #0  "function-inline-cost"="75" !prof !15 {
+  ret void
+}
+
+declare void @extern_f()
+attributes #0 = { cold }
+
+!15 = !{!"function_entry_count", i64 0}
+!16 = !{!"function_entry_count", i64 1000000}
+
+!llvm.module.flags = !{!1}
+!1 = !{i32 1, !"ProfileSummary", !2}
+!2 = !{!3, !4, !5, !6, !7, !8, !9, !10}
+!3 = !{!"ProfileFormat", !"InstrProf"}
+!4 = !{!"TotalCount", i64 1000000}
+!5 = !{!"MaxCount", i64 1000000}
+!6 = !{!"MaxInternalCount", i64 1}
+!7 = !{!"MaxFunctionCount", i64 1000000}
+!8 = !{!"NumCounts", i64 100}
+!9 = !{!"NumFunctions", i64 100}
+!10 = !{!"DetailedSummary", !11}
+!11 = !{!12, !13, !14}
+!12 = !{i32 10000, i64 100, i32 1}
+!13 = !{i32 990000, i64 100, i32 1}
+!14 = !{i32 999999, i64 1, i32 2}