Re-Reland [X86] Respect code models more when determining if a global reference can fit in 32 bits (llvm#75386)

For non-GlobalValue references, the small and medium code models can use
32 bit constants.

For GlobalValue references, use TargetMachine::isLargeGlobalObject().
Look through aliases for determining if a GlobalValue is small or large.
Even the large code model can reference small objects with 32 bit
constants as long as we're in no-pic mode, or if the reference is offset
from the GOT.

Original commit broke the build...

First reland broke large PIC builds referencing small data since it was using GOTOFF as a 32-bit constant.

Author: aeubanks

llvm/include/llvm/Target/TargetMachine.h

@@ -239,7 +239,7 @@ class TargetMachine {
   void setCodeModel(CodeModel::Model CM) { CMModel = CM; }
 
   void setLargeDataThreshold(uint64_t LDT) { LargeDataThreshold = LDT; }
-  bool isLargeGlobalObject(const GlobalObject *GO) const;
+  bool isLargeGlobalValue(const GlobalValue *GV) const;
 
   bool isPositionIndependent() const;
 

llvm/lib/CodeGen/TargetLoweringObjectFileImpl.cpp

@@ -649,7 +649,7 @@ getELFSectionNameForGlobal(const GlobalObject *GO, SectionKind Kind,
     Name = ".rodata.cst";
     Name += utostr(EntrySize);
   } else {
-    Name = getSectionPrefixForGlobal(Kind, TM.isLargeGlobalObject(GO));
+    Name = getSectionPrefixForGlobal(Kind, TM.isLargeGlobalValue(GO));
   }
 
   bool HasPrefix = false;
@@ -769,7 +769,7 @@ getGlobalObjectInfo(const GlobalObject *GO, const TargetMachine &TM) {
     Group = C->getName();
     IsComdat = C->getSelectionKind() == Comdat::Any;
   }
-  if (TM.isLargeGlobalObject(GO))
+  if (TM.isLargeGlobalValue(GO))
     Flags |= ELF::SHF_X86_64_LARGE;
   return {Group, IsComdat, Flags};
 }

llvm/lib/Target/TargetMachine.cpp

@@ -39,10 +39,16 @@ TargetMachine::TargetMachine(const Target &T, StringRef DataLayoutString,
 
 TargetMachine::~TargetMachine() = default;
 
-bool TargetMachine::isLargeGlobalObject(const GlobalObject *GO) const {
+bool TargetMachine::isLargeGlobalValue(const GlobalValue *GVal) const {
   if (getTargetTriple().getArch() != Triple::x86_64)
     return false;
 
+  auto *GO = GVal->getAliaseeObject();
+
+  // Be conservative if we can't find an underlying GlobalObject.
+  if (!GO)
+    return true;
+
   auto *GV = dyn_cast<GlobalVariable>(GO);
 
   // Functions/GlobalIFuncs are only large under the large code model.

llvm/lib/Target/X86/X86FastISel.cpp

@@ -716,10 +716,8 @@ bool X86FastISel::handleConstantAddresses(const Value *V, X86AddressMode &AM) {
       return false;
 
     // Can't handle large objects yet.
-    if (auto *GO = dyn_cast<GlobalObject>(GV)) {
-      if (TM.isLargeGlobalObject(GO))
-        return false;
-    }
+    if (TM.isLargeGlobalValue(GV))
+      return false;
 
     // Can't handle TLS yet.
     if (GV->isThreadLocal())
@@ -3849,7 +3847,7 @@ unsigned X86FastISel::X86MaterializeGV(const GlobalValue *GV, MVT VT) {
   if (TM.getCodeModel() != CodeModel::Small &&
       TM.getCodeModel() != CodeModel::Medium)
     return 0;
-  if (!isa<GlobalObject>(GV) || TM.isLargeGlobalObject(cast<GlobalObject>(GV)))
+  if (TM.isLargeGlobalValue(GV))
     return 0;
 
   // Materialize addresses with LEA/MOV instructions.

llvm/lib/Target/X86/X86ISelDAGToDAG.cpp

@@ -2927,6 +2927,13 @@ bool X86DAGToDAGISel::selectAddr(SDNode *Parent, SDValue N, SDValue &Base,
 }
 
 bool X86DAGToDAGISel::selectMOV64Imm32(SDValue N, SDValue &Imm) {
+  // Cannot use 32 bit constants to reference objects in kernel code model.
+  // Cannot use 32 bit constants to reference objects in large PIC mode since
+  // GOTOFF is 64 bits.
+  if (TM.getCodeModel() == CodeModel::Kernel ||
+      (TM.getCodeModel() == CodeModel::Large && TM.isPositionIndependent()))
+    return false;
+
   // In static codegen with small code model, we can get the address of a label
   // into a register with 'movl'
   if (N->getOpcode() != X86ISD::Wrapper)
@@ -2940,15 +2947,18 @@ bool X86DAGToDAGISel::selectMOV64Imm32(SDValue N, SDValue &Imm) {
     return false;
 
   Imm = N;
-  if (N->getOpcode() != ISD::TargetGlobalAddress)
-    return TM.getCodeModel() == CodeModel::Small;
+  // Small/medium code model can reference non-TargetGlobalAddress objects with
+  // 32 bit constants.
+  if (N->getOpcode() != ISD::TargetGlobalAddress) {
+    return TM.getCodeModel() == CodeModel::Small ||
+           TM.getCodeModel() == CodeModel::Medium;
+  }
 
-  std::optional<ConstantRange> CR =
-      cast<GlobalAddressSDNode>(N)->getGlobal()->getAbsoluteSymbolRange();
-  if (!CR)
-    return TM.getCodeModel() == CodeModel::Small;
+  const GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
+  if (std::optional<ConstantRange> CR = GV->getAbsoluteSymbolRange())
+    return CR->getUnsignedMax().ult(1ull << 32);
 
-  return CR->getUnsignedMax().ult(1ull << 32);
+  return !TM.isLargeGlobalValue(GV);
 }
 
 bool X86DAGToDAGISel::selectLEA64_32Addr(SDValue N, SDValue &Base,

llvm/lib/Target/X86/X86Subtarget.cpp

@@ -86,16 +86,16 @@ X86Subtarget::classifyLocalReference(const GlobalValue *GV) const {
       CodeModel::Model CM = TM.getCodeModel();
       assert(CM != CodeModel::Tiny &&
              "Tiny codesize model not supported on X86");
-      // In the large code model, even referencing a global under the large data
-      // threshold which is considered "small", we need to use GOTOFF.
+      // In the large code model, all text is far from any global data, so we
+      // use GOTOFF.
       if (CM == CodeModel::Large)
         return X86II::MO_GOTOFF;
-      // Large objects use GOTOFF, otherwise use RIP-rel access.
-      if (auto *GO = dyn_cast_or_null<GlobalObject>(GV))
-        return TM.isLargeGlobalObject(GO) ? X86II::MO_GOTOFF
-                                          : X86II::MO_NO_FLAG;
-      // For non-GlobalObjects, the small and medium code models treat them as
-      // accessible with a RIP-rel access.
+      // Large GlobalValues use GOTOFF, otherwise use RIP-rel access.
+      if (GV)
+        return TM.isLargeGlobalValue(GV) ? X86II::MO_GOTOFF : X86II::MO_NO_FLAG;
+      // GV == nullptr is for all other non-GlobalValue global data like the
+      // constant pool, jump tables, labels, etc. The small and medium code
+      // models treat these as accessible with a RIP-rel access.
       return X86II::MO_NO_FLAG;
     }