Add a new fixed metadata node to carry the xcoff pgo association.

Instead of decomposing the initializer for a __profd profiling data
symbol to get its associated __profc symbol attach a metadata node
tying the 2 respective globals together.

Author: mandlebug

llvm/include/llvm/IR/FixedMetadataKinds.def

@@ -54,3 +54,4 @@ LLVM_FIXED_MD_KIND(MD_coro_outside_frame, "coro.outside.frame", 39)
 LLVM_FIXED_MD_KIND(MD_mmra, "mmra", 40)
 LLVM_FIXED_MD_KIND(MD_noalias_addrspace, "noalias.addrspace", 41)
 LLVM_FIXED_MD_KIND(MD_callee_type, "callee_type", 42)
+LLVM_FIXED_MD_KIND(MD_pgo_associated, "pgo.associated", 43)

llvm/lib/Target/PowerPC/PPCAsmPrinter.cpp

@@ -2813,46 +2813,17 @@ void PPCAIXAsmPrinter::emitGlobalVariableHelper(const GlobalVariable *GV) {
   // associated profiling counters.
   if (TM.getFunctionSections() &&
       Csect->getName().starts_with("__llvm_prf_data.")) {
-    // Unraveling the initializer to find the related counters variable. The
-    // initializer is a structure whose third member is a subtract expression
-    // between the counters label and the label for the start of this structure.
-    // Use the subtract expression to get the GlobalValue for the counters
-    // global.
-    assert(GV->hasInitializer() &&
-           "profiling data symbol must have an initializer");
-    assert(isa<ConstantStruct>(GV->getInitializer()) &&
-           "expect the initializer for a profiling data symbol to be a struct");
-    const ConstantStruct *Initializer =
-        cast<ConstantStruct>(GV->getInitializer());
-
-    // The initializer structure is: { i64, i64, i32, ptr, ptr, i32, [4 x i16] }
-    // and the reference to the global variable for the counters is in the
-    // first i32 member.
-    const Constant *Member = Initializer->getAggregateElement(2);
-    assert(Member && "profiling data symbol has more then 3 elements");
-
-    // Want to decompose a constant expression of the form:
-    // sub (ptrtoint (ptr @__profc_sym), ptrtoint (ptr @__profd_sym))
-    // to get the GlobalVariable for the '@__profc_sym` symbol.
-    assert(isa<ConstantExpr>(Member) &&
-           "expected member initializer is a constant expression.");
-    const ConstantExpr *CExpr = cast<ConstantExpr>(Member);
-    assert(CExpr->getOpcode() == Instruction::Sub &&
-           "expected member intializer is a sub expression.");
-
-    Value *V1 = CExpr->getOperand(0);
-    assert(V1 && isa<ConstantExpr>(V1) &&
-           "expected sub expression operand to be constant expr.");
-    ConstantExpr *PointerToIntExpr = cast<ConstantExpr>(V1);
-    assert(PointerToIntExpr->isCast() && "unexpected operand type.");
-
-    Value *PointerToIntOperand = PointerToIntExpr->getOperand(0);
-    assert(isa<GlobalVariable>(PointerToIntOperand) &&
-           "expected global variable of profc symbol");
-
-    const GlobalVariable *ProfCGV = cast<GlobalVariable>(PointerToIntOperand);
+    MDNode *MD = GV->getMetadata(LLVMContext::MD_pgo_associated);
+
+    assert(MD &&
+           "profiling data symbol must have an associated counter symbol");
+
+    const ValueAsMetadata *VAM = cast<ValueAsMetadata>(MD->getOperand(0).get());
+    const GlobalVariable *ProfCGV = cast<GlobalVariable>(VAM->getValue());
+
     // Map the global variable to its CSECT.
     SectionKind ProfCKind = getObjFileLowering().getKindForGlobal(GV, TM);
+
     MCSectionXCOFF *ProfCCsect = cast<MCSectionXCOFF>(
         getObjFileLowering().SectionForGlobal(ProfCGV, ProfCKind, TM));
 

llvm/lib/Transforms/Instrumentation/InstrProfiling.cpp

@@ -1875,6 +1875,12 @@ void InstrLowerer::createDataVariable(InstrProfCntrInstBase *Inc) {
   Data->setAlignment(Align(INSTR_PROF_DATA_ALIGNMENT));
   maybeSetComdat(Data, Fn, CntsVarName);
 
+  if (TT.isOSBinFormatXCOFF()) {
+    Data->setMetadata(
+        LLVMContext::MD_pgo_associated,
+        MDNode::get(M.getContext(), ValueAsMetadata::get(PD.RegionCounters)));
+  }
+
   PD.DataVar = Data;
 
   // Mark the data variable as used so that it isn't stripped out.

llvm/test/CodeGen/PowerPC/aix-pgo-function-sections.ll

@@ -7,9 +7,9 @@
 @i = external local_unnamed_addr global i32, align 4
 @__llvm_profile_raw_version = weak hidden local_unnamed_addr constant i64 72057594037927944
 @__profc_func1 = private global [1 x i64] zeroinitializer, section "__llvm_prf_cnts", align 8
-@__profd_func1 = private global { i64, i64, i32, ptr, ptr, i32, [4 x i16] } { i64 -2545542355363006406, i64 742261418966908927, i32 sub (i32 ptrtoint (ptr @__profc_func1 to i32), i32 ptrtoint (ptr @__profd_func1 to i32)), ptr @func1.local, ptr null, i32 1, [4 x i16] zeroinitializer }, section "__llvm_prf_data", align 8
+@__profd_func1 = private global { i64, i64, i32, ptr, ptr, i32, [4 x i16] } { i64 -2545542355363006406, i64 742261418966908927, i32 sub (i32 ptrtoint (ptr @__profc_func1 to i32), i32 ptrtoint (ptr @__profd_func1 to i32)), ptr @func1.local, ptr null, i32 1, [4 x i16] zeroinitializer }, section "__llvm_prf_data", align 8, !pgo.associated !0
 @__profc_func2 = private global [1 x i64] zeroinitializer, section "__llvm_prf_cnts", align 8
-@__profd_func2 = private global { i64, i64, i32, ptr, ptr, i32, [4 x i16] } { i64 -4377547752858689819, i64 742261418966908927, i32 sub (i32 ptrtoint (ptr @__profc_func2 to i32), i32 ptrtoint (ptr @__profd_func2 to i32)), ptr @func2.local, ptr null, i32 1, [4 x i16] zeroinitializer }, section "__llvm_prf_data", align 8
+@__profd_func2 = private global { i64, i64, i32, ptr, ptr, i32, [4 x i16] } { i64 -4377547752858689819, i64 742261418966908927, i32 sub (i32 ptrtoint (ptr @__profc_func2 to i32), i32 ptrtoint (ptr @__profd_func2 to i32)), ptr @func2.local, ptr null, i32 1, [4 x i16] zeroinitializer }, section "__llvm_prf_data", align 8, !pgo.associated !1
 @__llvm_prf_nm = private constant [13 x i8] c"\0B\00func1\01func2", section "__llvm_prf_names", align 1
 @__llvm_profile_filename = weak hidden local_unnamed_addr constant [19 x i8] c"default_%m.profraw\00"
 @llvm.used = appending global [3 x ptr] [ptr @__llvm_prf_nm, ptr @__profd_func1, ptr @__profd_func2], section "llvm.metadata"
@@ -38,6 +38,9 @@ entry:
 
 declare i32 @external_func(i32 noundef)
 
+!0 = !{ptr @__profc_func1}
+!1 = !{ptr @__profc_func2}
+
 ; CHECK-DAG:        .csect __llvm_prf_cnts.__profc_func1[RW]
 ; CHECK-DAG:        .csect __llvm_prf_data.__profd_func1[RW]
 ; CHECK-DAG:        .csect __llvm_prf_cnts.__profc_func2[RW]

llvm/test/LTO/PowerPC/pgo-function-sections-aix.ll

@@ -16,11 +16,11 @@ target triple = "powerpc-ibm-aix7.2.0.0"
 
 @__llvm_profile_raw_version = weak hidden constant i64 72057594037927944
 @__profc_func1 = private global [2 x i64] zeroinitializer, section "__llvm_prf_cnts", align 8
-@__profd_func1 = private global { i64, i64, i32, ptr, ptr, i32, [4 x i16] } { i64 -2545542355363006406, i64 146835647075900052, i32 sub (i32 ptrtoint (ptr @__profc_func1 to i32), i32 ptrtoint (ptr @__profd_func1 to i32)), ptr @func1.local, ptr null, i32 2, [4 x i16] zeroinitializer }, section "__llvm_prf_data", align 8
+@__profd_func1 = private global { i64, i64, i32, ptr, ptr, i32, [4 x i16] } { i64 -2545542355363006406, i64 146835647075900052, i32 sub (i32 ptrtoint (ptr @__profc_func1 to i32), i32 ptrtoint (ptr @__profd_func1 to i32)), ptr @func1.local, ptr null, i32 2, [4 x i16] zeroinitializer }, section "__llvm_prf_data", align 8, !pgo.associated !0
 @__profc_file1.c_internal_func = private global [1 x i64] zeroinitializer, section "__llvm_prf_cnts", align 8
-@__profd_file1.c_internal_func = private global { i64, i64, i32, ptr, ptr, i32, [4 x i16] } { i64 2905054957054668920, i64 742261418966908927, i32 sub (i32 ptrtoint (ptr @__profc_file1.c_internal_func to i32), i32 ptrtoint(ptr @__profd_file1.c_internal_func to i32)), ptr @internal_func, ptr null, i32 1, [4 x i16] zeroinitializer }, section "__llvm_prf_data", align 8
+@__profd_file1.c_internal_func = private global { i64, i64, i32, ptr, ptr, i32, [4 x i16] } { i64 2905054957054668920, i64 742261418966908927, i32 sub (i32 ptrtoint (ptr @__profc_file1.c_internal_func to i32), i32 ptrtoint(ptr @__profd_file1.c_internal_func to i32)), ptr @internal_func, ptr null, i32 1, [4 x i16] zeroinitializer }, section "__llvm_prf_data", align 8, !pgo.associated !1
 @__profc_escape1 = private global [1 x i64] zeroinitializer, section "__llvm_prf_cnts", align 8
-@__profd_escape1 = private global { i64, i64, i32, ptr, ptr, i32, [4 x i16] } { i64 3473293639883741762, i64 742261418966908927, i32 sub (i32 ptrtoint (ptr @__profc_escape1 to i32), i32 ptrtoint (ptr @__profd_escape1 to i32)), ptr @escape1.local, ptr null, i32 1, [4 x i16] zeroinitializer }, section "__llvm_prf_data", align 8
+@__profd_escape1 = private global { i64, i64, i32, ptr, ptr, i32, [4 x i16] } { i64 3473293639883741762, i64 742261418966908927, i32 sub (i32 ptrtoint (ptr @__profc_escape1 to i32), i32 ptrtoint (ptr @__profd_escape1 to i32)), ptr @escape1.local, ptr null, i32 1, [4 x i16] zeroinitializer }, section "__llvm_prf_data", align 8, !pgo.associated !2
 @__llvm_prf_nm = private constant [37 x i8] c"#\00func1\01file1.c:internal_func\01escape1", section "__llvm_prf_names", align 1
 @llvm.used = appending global [4 x ptr] [ptr @__profd_func1, ptr @__profd_file1.c_internal_func, ptr @__profd_escape1, ptr @__llvm_prf_nm], section "llvm.metadata"
 @__llvm_profile_filename = weak hidden constant [19 x i8] c"default_%m.profraw\00"
@@ -87,17 +87,21 @@ declare void @llvm.instrprof.increment(ptr, i64, i32, i32) #1
 attributes #0 = { noinline nounwind optnone }
 attributes #1 = { nounwind }
 
+!0 = !{ptr @__profc_func1}
+!1 = !{ptr @__profc_file1.c_internal_func}
+!2 = !{ptr @__profc_escape1}
+
 ;--- f2.ll
 target datalayout = "E-m:a-p:32:32-i64:64-n32"
 target triple = "powerpc-ibm-aix7.2.0.0"
 
 @__llvm_profile_raw_version = weak hidden constant i64 72057594037927944
 @__profc_func2 = private global [2 x i64] zeroinitializer, section "__llvm_prf_cnts", align 8
-@__profd_func2 = private global { i64, i64, i32, ptr, ptr, i32, [4 x i16] } { i64 -4377547752858689819, i64 146835647075900052, i32 sub (i32 ptrtoint (ptr @__profc_func2 to i32), i32 ptrtoint (ptr @__profd_func2 to i32)), ptr @func2.local, ptr null, i32 2, [4 x i16] zeroinitializer }, section "__llvm_prf_data", align 8
+@__profd_func2 = private global { i64, i64, i32, ptr, ptr, i32, [4 x i16] } { i64 -4377547752858689819, i64 146835647075900052, i32 sub (i32 ptrtoint (ptr @__profc_func2 to i32), i32 ptrtoint (ptr @__profd_func2 to i32)), ptr @func2.local, ptr null, i32 2, [4 x i16] zeroinitializer }, section "__llvm_prf_data", align 8, !pgo.associated !0
 @__profc_file2.c_internal_func = private global [1 x i64] zeroinitializer, section "__llvm_prf_cnts", align 8
-@__profd_file2.c_internal_func = private global { i64, i64, i32, ptr, ptr, i32, [4 x i16] } { i64 4899437111831460578, i64 742261418966908927, i32 sub (i32 ptrtoint (ptr @__profc_file2.c_internal_func to i32), i32 ptrtoint (ptr @__profd_file2.c_internal_func to i32)), ptr @internal_func, ptr null, i32 1, [4 x i16] zeroinitializer }, section "__llvm_prf_data", align 8
+@__profd_file2.c_internal_func = private global { i64, i64, i32, ptr, ptr, i32, [4 x i16] } { i64 4899437111831460578, i64 742261418966908927, i32 sub (i32 ptrtoint (ptr @__profc_file2.c_internal_func to i32), i32 ptrtoint (ptr @__profd_file2.c_internal_func to i32)), ptr @internal_func, ptr null, i32 1, [4 x i16] zeroinitializer }, section "__llvm_prf_data", align 8, !pgo.associated !1
 @__profc_escape2 = private global [1 x i64] zeroinitializer, section "__llvm_prf_cnts", align 8
-@__profd_escape2 = private global { i64, i64, i32, ptr, ptr, i32, [4 x i16] } { i64 3333263850724280696, i64 742261418966908927, i32 sub (i32 ptrtoint (ptr @__profc_escape2 to i32), i32 ptrtoint (ptr @__profd_escape2 to i32)), ptr @escape2.local, ptr null, i32 1, [4 x i16] zeroinitializer }, section "__llvm_prf_data", align 8
+@__profd_escape2 = private global { i64, i64, i32, ptr, ptr, i32, [4 x i16] } { i64 3333263850724280696, i64 742261418966908927, i32 sub (i32 ptrtoint (ptr @__profc_escape2 to i32), i32 ptrtoint (ptr @__profd_escape2 to i32)), ptr @escape2.local, ptr null, i32 1, [4 x i16] zeroinitializer }, section "__llvm_prf_data", align 8, !pgo.associated !2
 @__llvm_prf_nm = private constant [37 x i8] c"#\00func2\01file2.c:internal_func\01escape2", section "__llvm_prf_names", align 1
 @llvm.used = appending global [4 x ptr] [ptr @__profd_func2, ptr @__profd_file2.c_internal_func, ptr @__profd_escape2, ptr @__llvm_prf_nm], section "llvm.metadata"
 @__llvm_profile_filename = weak hidden constant [19 x i8] c"default_%m.profraw\00"
@@ -166,6 +170,10 @@ declare void @llvm.instrprof.increment(ptr, i64, i32, i32) #1
 attributes #0 = { noinline nounwind optnone }
 attributes #1 = { nounwind }
 
+!0 = !{ptr @__profc_func2}
+!1 = !{ptr @__profc_file2.c_internal_func}
+!2 = !{ptr @__profc_escape2}
+
 ; CHECK-DAG:        .csect __llvm_prf_cnts.__profc_func1[RW]
 ; CHECK-DAG:        .csect __llvm_prf_data.__profd_func1[RW]
 ; CHECK-DAG:        .csect __llvm_prf_cnts.__profc_file1.c_internal_func[RW]