[PGO] Make the PGO instrumentation insert point after alloca

We're changing PGO instrumentation to insert the intrinsic after
alloca instructions. For sampled instrumentation, a conditional
check is placed before the intrinsic. If this intrinsic comes
before an alloca, the alloca (whose size might be unknown due to
Phi node) becomes conditional, resulting in inefficient code.
We have seen some stack overflows due to this.

This patch guarantees the intrinsic is always after the alloca.

Author: xur-llvm

llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp

@@ -855,7 +855,7 @@ BasicBlock *FuncPGOInstrumentation<Edge, BBInfo>::getInstrBB(Edge *E) {
   auto canInstrument = [](BasicBlock *BB) -> BasicBlock * {
     // There are basic blocks (such as catchswitch) cannot be instrumented.
     // If the returned first insertion point is the end of BB, skip this BB.
-    if (BB->getFirstInsertionPt() == BB->end())
+    if (BB->getFirstNonPHIOrDbgOrAlloca() == BB->end())
       return nullptr;
     return BB;
   };
@@ -952,7 +952,7 @@ void FunctionInstrumenter::instrument() {
       Name, PointerType::get(M.getContext(), 0));
   if (PGOFunctionEntryCoverage) {
     auto &EntryBB = F.getEntryBlock();
-    IRBuilder<> Builder(&EntryBB, EntryBB.getFirstInsertionPt());
+    IRBuilder<> Builder(&EntryBB, EntryBB.getFirstNonPHIOrDbgOrAlloca());
     // llvm.instrprof.cover(i8* <name>, i64 <hash>, i32 <num-counters>,
     //                      i32 <index>)
     Builder.CreateIntrinsic(
@@ -1010,7 +1010,7 @@ void FunctionInstrumenter::instrument() {
   if (PGOTemporalInstrumentation) {
     NumCounters += PGOBlockCoverage ? 8 : 1;
     auto &EntryBB = F.getEntryBlock();
-    IRBuilder<> Builder(&EntryBB, EntryBB.getFirstInsertionPt());
+    IRBuilder<> Builder(&EntryBB, EntryBB.getFirstNonPHIOrDbgOrAlloca());
     // llvm.instrprof.timestamp(i8* <name>, i64 <hash>, i32 <num-counters>,
     //                          i32 <index>)
     Builder.CreateIntrinsic(Intrinsic::instrprof_timestamp,
@@ -1021,7 +1021,7 @@ void FunctionInstrumenter::instrument() {
   }
 
   for (auto *InstrBB : InstrumentBBs) {
-    IRBuilder<> Builder(InstrBB, InstrBB->getFirstInsertionPt());
+    IRBuilder<> Builder(InstrBB, InstrBB->getFirstNonPHIOrDbgOrAlloca());
     assert(Builder.GetInsertPoint() != InstrBB->end() &&
            "Cannot get the Instrumentation point");
     // llvm.instrprof.increment(i8* <name>, i64 <hash>, i32 <num-counters>,

llvm/test/Transforms/PGOProfile/entry_alloca.ll

@@ -0,0 +1,44 @@
+; Note: Make sure that instrumention intrinsic is after entry alloca.
+; RUN: opt < %s -passes=pgo-instr-gen -S | FileCheck %s 
+; RUN: opt < %s -passes=pgo-instr-gen,instrprof -sampled-instrumentation -S | FileCheck %s --check-prefixes=SAMPLE
+
+%struct.A = type { i32, [0 x i32] }
+
+; CHECK-LABEL: @foo()
+; CHECK-NEXT:   %1 = alloca %struct.A
+; CHECK-NEXT:   call void @llvm.instrprof.increment(ptr @__profn_foo
+; CHECK-NEXT:   call void @bar(ptr
+
+; SAMPLE: @foo()
+; SAMPLE-NEXT:  %1 = alloca %struct.A, align 4
+; SAMPLE-NEXT:  %[[v:[0-9]+]] = load i16, ptr @__llvm_profile_sampling
+; SAMPLE-NEXT:  {{.*}} = icmp ule i16 %[[v]], 199
+
+define dso_local i32 @foo() {
+  %1 = alloca %struct.A, align 4
+  call void @bar(ptr noundef nonnull %1) #3
+  %2 = load i32, ptr %1, align 4
+  %3 = icmp sgt i32 %2, 0
+  br i1 %3, label %4, label %15
+
+4:
+  %5 = getelementptr inbounds i8, ptr %1, i64 4
+  %6 = zext nneg i32 %2 to i64
+  br label %7
+
+7:
+  %8 = phi i64 [ 0, %4 ], [ %13, %7 ]
+  %9 = phi i32 [ 0, %4 ], [ %12, %7 ]
+  %10 = getelementptr inbounds [0 x i32], ptr %5, i64 0, i64 %8
+  %11 = load i32, ptr %10, align 4
+  %12 = add nsw i32 %11, %9
+  %13 = add nuw nsw i64 %8, 1
+  %14 = icmp eq i64 %13, %6
+  br i1 %14, label %15, label %7
+
+15:
+  %16 = phi i32 [ 0, %0 ], [ %12, %7 ]
+  ret i32 %16
+}
+
+declare void @bar(ptr noundef)

llvm/test/Transforms/PGOProfile/split-indirectbr-critical-edges.ll

@@ -42,6 +42,7 @@ if.end:                                           ; preds = %if.end.preheader, %
 ;; The edge will not be profiled.
 ; CHECK-LABEL: @cannot_split(
 ; CHECK-NEXT:  entry:
+; CHECK-NEXT:    %targets = alloca <2 x ptr>, align 16
 ; CHECK-NEXT:    call void @llvm.instrprof.increment
 ; CHECK: indirect:
 ; CHECK-NOT:     call void @llvm.instrprof.increment