[PowerPC][GlobalMerge] Reduce TOC usage by merging internal and private global data

llvm/include/llvm/CodeGen/GlobalMerge.h

@@ -28,6 +28,8 @@ struct GlobalMergeOptions {
   bool MergeConst = false;
   /// Whether we should merge global variables that have external linkage.
   bool MergeExternal = true;
+  /// Whether we should merge constant global variables.
+  bool MergeConstantGlobals = false;
   /// Whether we should try to optimize for size only.
   /// Currently, this applies a dead simple heuristic: only consider globals
   /// used in minsize functions for merging.

llvm/include/llvm/CodeGen/Passes.h

@@ -476,7 +476,8 @@ namespace llvm {
   ///
   Pass *createGlobalMergePass(const TargetMachine *TM, unsigned MaximalOffset,
                               bool OnlyOptimizeForSize = false,
-                              bool MergeExternalByDefault = false);
+                              bool MergeExternalByDefault = false,
+                              bool MergeConstantByDefault = false);
 
   /// This pass splits the stack into a safe stack and an unsafe stack to
   /// protect against stack-based overflow vulnerabilities.

llvm/lib/CodeGen/GlobalMerge.cpp

@@ -196,11 +196,13 @@ class GlobalMerge : public FunctionPass {
   }
 
   explicit GlobalMerge(const TargetMachine *TM, unsigned MaximalOffset,
-                       bool OnlyOptimizeForSize, bool MergeExternalGlobals)
+                       bool OnlyOptimizeForSize, bool MergeExternalGlobals,
+                       bool MergeConstantGlobals)
       : FunctionPass(ID), TM(TM) {
     Opt.MaxOffset = MaximalOffset;
     Opt.SizeOnly = OnlyOptimizeForSize;
     Opt.MergeExternal = MergeExternalGlobals;
+    Opt.MergeConstantGlobals = MergeConstantGlobals;
     initializeGlobalMergePass(*PassRegistry::getPassRegistry());
   }
 
@@ -475,7 +477,8 @@ bool GlobalMergeImpl::doMerge(const SmallVectorImpl<GlobalVariable *> &Globals,
   auto &DL = M.getDataLayout();
 
   LLVM_DEBUG(dbgs() << " Trying to merge set, starts with #"
-                    << GlobalSet.find_first() << "\n");
+                    << GlobalSet.find_first() << ", total of " << Globals.size()
+                    << "\n");
 
   bool Changed = false;
   ssize_t i = GlobalSet.find_first();
@@ -551,6 +554,8 @@ bool GlobalMergeImpl::doMerge(const SmallVectorImpl<GlobalVariable *> &Globals,
     MergedGV->setAlignment(MaxAlign);
     MergedGV->setSection(Globals[i]->getSection());
 
+    LLVM_DEBUG(dbgs() << "MergedGV:  " << *MergedGV << "\n");
+
     const StructLayout *MergedLayout = DL.getStructLayout(MergedTy);
     for (ssize_t k = i, idx = 0; k != j; k = GlobalSet.find_next(k), ++idx) {
       GlobalValue::LinkageTypes Linkage = Globals[k]->getLinkage();
@@ -700,6 +705,11 @@ bool GlobalMergeImpl::run(Module &M) {
       else
         Globals[{AddressSpace, Section}].push_back(&GV);
     }
+    LLVM_DEBUG(dbgs() << "GV "
+                      << ((DL.getTypeAllocSize(Ty) < Opt.MaxOffset)
+                              ? "to merge: "
+                              : "not to merge: ")
+                      << GV << "\n");
   }
 
   for (auto &P : Globals)
@@ -710,7 +720,7 @@ bool GlobalMergeImpl::run(Module &M) {
     if (P.second.size() > 1)
       Changed |= doMerge(P.second, M, false, P.first.first);
 
-  if (EnableGlobalMergeOnConst)
+  if (Opt.MergeConstantGlobals)
     for (auto &P : ConstGlobals)
       if (P.second.size() > 1)
         Changed |= doMerge(P.second, M, true, P.first.first);
@@ -720,8 +730,11 @@ bool GlobalMergeImpl::run(Module &M) {
 
 Pass *llvm::createGlobalMergePass(const TargetMachine *TM, unsigned Offset,
                                   bool OnlyOptimizeForSize,
-                                  bool MergeExternalByDefault) {
+                                  bool MergeExternalByDefault,
+                                  bool MergeConstantByDefault) {
   bool MergeExternal = (EnableGlobalMergeOnExternal == cl::BOU_UNSET) ?
     MergeExternalByDefault : (EnableGlobalMergeOnExternal == cl::BOU_TRUE);
-  return new GlobalMerge(TM, Offset, OnlyOptimizeForSize, MergeExternal);
+  bool MergeConstant = EnableGlobalMergeOnConst || MergeConstantByDefault;
+  return new GlobalMerge(TM, Offset, OnlyOptimizeForSize, MergeExternal,
+                         MergeConstant);
 }

llvm/lib/Target/PowerPC/PPCTargetMachine.cpp

@@ -111,6 +111,15 @@ static cl::opt<bool> EnablePPCGenScalarMASSEntries(
              "(scalar) entries"),
     cl::Hidden);
 
+static cl::opt<bool>
+    EnableGlobalMerge("ppc-global-merge", cl::Hidden, cl::init(false),
+                      cl::desc("Enable the global merge pass"));
+
+static cl::opt<unsigned>
+    GlobalMergeMaxOffset("ppc-global-merge-max-offset", cl::Hidden,
+                         cl::init(0x7fff),
+                         cl::desc("Maximum global merge offset"));
+
 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializePowerPCTarget() {
   // Register the targets
   RegisterTargetMachine<PPCTargetMachine> A(getThePPC32Target());
@@ -491,6 +500,10 @@ void PPCPassConfig::addIRPasses() {
 }
 
 bool PPCPassConfig::addPreISel() {
+  if (EnableGlobalMerge)
+    addPass(
+        createGlobalMergePass(TM, GlobalMergeMaxOffset, false, false, true));
+
   if (MergeStringPool && getOptLevel() != CodeGenOptLevel::None)
     addPass(createPPCMergeStringPoolPass());
 

llvm/test/CodeGen/PowerPC/merge-private.ll

@@ -1,15 +1,15 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
 ; RUN: llc -verify-machineinstrs -mtriple powerpc64-ibm-aix-xcoff -mcpu=pwr8 \
-; RUN:     -ppc-asm-full-reg-names < %s | FileCheck %s \
+; RUN:     -ppc-asm-full-reg-names -ppc-global-merge=true < %s | FileCheck %s \
 ; RUN:     --check-prefix=AIX64
 ; RUN: llc -verify-machineinstrs -mtriple powerpc-ibm-aix-xcoff -mcpu=pwr8 \
-; RUN:     -ppc-asm-full-reg-names < %s | FileCheck %s \
+; RUN:     -ppc-asm-full-reg-names -ppc-global-merge=true < %s | FileCheck %s \
 ; RUN:     --check-prefix=AIX32
 ; RUN: llc -verify-machineinstrs -mtriple powerpc64le-unknown-linux -mcpu=pwr8 \
-; RUN:     -ppc-asm-full-reg-names < %s | FileCheck %s \
+; RUN:     -ppc-asm-full-reg-names -ppc-global-merge=true < %s | FileCheck %s \
 ; RUN:     --check-prefix=LINUX64LE
 ; RUN: llc -verify-machineinstrs -mtriple powerpc64-unknown-linux -mcpu=pwr8 \
-; RUN:     -ppc-asm-full-reg-names < %s | FileCheck %s \
+; RUN:     -ppc-asm-full-reg-names -ppc-global-merge=true < %s | FileCheck %s \
 ; RUN:     --check-prefix=LINUX64BE
 
 @.str = private unnamed_addr constant [15 x i8] c"Private global\00", align 1
@@ -24,7 +24,7 @@ define dso_local void @print_func() {
 ; AIX64-NEXT:    stdu r1, -128(r1)
 ; AIX64-NEXT:    std r0, 144(r1)
 ; AIX64-NEXT:    std r31, 120(r1) # 8-byte Folded Spill
-; AIX64-NEXT:    ld r31, L..C0(r2) # @__ModuleStringPool
+; AIX64-NEXT:    ld r31, L..C0(r2) # @_MergedGlobals
 ; AIX64-NEXT:    mr r3, r31
 ; AIX64-NEXT:    bl .puts[PR]
 ; AIX64-NEXT:    nop
@@ -43,7 +43,7 @@ define dso_local void @print_func() {
 ; AIX32-NEXT:    stwu r1, -64(r1)
 ; AIX32-NEXT:    stw r0, 72(r1)
 ; AIX32-NEXT:    stw r31, 60(r1) # 4-byte Folded Spill
-; AIX32-NEXT:    lwz r31, L..C0(r2) # @__ModuleStringPool
+; AIX32-NEXT:    lwz r31, L..C0(r2) # @_MergedGlobals
 ; AIX32-NEXT:    mr r3, r31
 ; AIX32-NEXT:    bl .puts[PR]
 ; AIX32-NEXT:    nop
@@ -64,9 +64,9 @@ define dso_local void @print_func() {
 ; LINUX64LE-NEXT:    .cfi_offset r30, -16
 ; LINUX64LE-NEXT:    std r30, -16(r1) # 8-byte Folded Spill
 ; LINUX64LE-NEXT:    stdu r1, -48(r1)
-; LINUX64LE-NEXT:    addis r3, r2, .L__ModuleStringPool@toc@ha
+; LINUX64LE-NEXT:    addis r3, r2, .L_MergedGlobals@toc@ha
 ; LINUX64LE-NEXT:    std r0, 64(r1)
-; LINUX64LE-NEXT:    addi r30, r3, .L__ModuleStringPool@toc@l
+; LINUX64LE-NEXT:    addi r30, r3, .L_MergedGlobals@toc@l
 ; LINUX64LE-NEXT:    mr r3, r30
 ; LINUX64LE-NEXT:    bl puts
 ; LINUX64LE-NEXT:    nop
@@ -87,9 +87,9 @@ define dso_local void @print_func() {
 ; LINUX64BE-NEXT:    .cfi_def_cfa_offset 128
 ; LINUX64BE-NEXT:    .cfi_offset lr, 16
 ; LINUX64BE-NEXT:    .cfi_offset r30, -16
-; LINUX64BE-NEXT:    addis r3, r2, .L__ModuleStringPool@toc@ha
+; LINUX64BE-NEXT:    addis r3, r2, .L_MergedGlobals@toc@ha
 ; LINUX64BE-NEXT:    std r30, 112(r1) # 8-byte Folded Spill
-; LINUX64BE-NEXT:    addi r30, r3, .L__ModuleStringPool@toc@l
+; LINUX64BE-NEXT:    addi r30, r3, .L_MergedGlobals@toc@l
 ; LINUX64BE-NEXT:    mr r3, r30
 ; LINUX64BE-NEXT:    bl puts
 ; LINUX64BE-NEXT:    nop