[PowerPC][AIX] Enable aix-small-local-dynamic-tls target attribute (#86641)

Following the aix-small-local-exec-tls target attribute, this patch adds
a target attribute for an AIX-specific option in llc that informs the
compiler that it can use a faster access sequence for the local-dynamic
TLS model (formally named aix-small-local-dynamic-tls) when TLS
variables are less than ~32KB in size.

The patch either produces an addi/la with a displacement off of module
handle (return value from .__tls_get_mod) when the address is
calculated, or it produces an addi/la followed by a load/store when the
address is calculated and used for further accesses.

---------

Co-authored-by: Amy Kwan <[email protected]>

Author: Felix (Ting Wang)

llvm/lib/Target/PowerPC/MCTargetDesc/PPCXCOFFObjectWriter.cpp

@@ -71,6 +71,8 @@ std::pair<uint8_t, uint8_t> PPCXCOFFObjectWriter::getRelocTypeAndSignSize(
       return {XCOFF::RelocationType::R_TOCL, SignAndSizeForHalf16};
     case MCSymbolRefExpr::VK_PPC_AIX_TLSLE:
       return {XCOFF::RelocationType::R_TLS_LE, SignAndSizeForHalf16};
+    case MCSymbolRefExpr::VK_PPC_AIX_TLSLD:
+      return {XCOFF::RelocationType::R_TLS_LD, SignAndSizeForHalf16};
     }
   } break;
   case PPC::fixup_ppc_half16ds:
@@ -86,6 +88,8 @@ std::pair<uint8_t, uint8_t> PPCXCOFFObjectWriter::getRelocTypeAndSignSize(
       return {XCOFF::RelocationType::R_TOCL, 15};
     case MCSymbolRefExpr::VK_PPC_AIX_TLSLE:
       return {XCOFF::RelocationType::R_TLS_LE, 15};
+    case MCSymbolRefExpr::VK_PPC_AIX_TLSLD:
+      return {XCOFF::RelocationType::R_TLS_LD, 15};
     }
   } break;
   case PPC::fixup_ppc_br24:

llvm/lib/Target/PowerPC/PPC.td

@@ -329,6 +329,15 @@ def FeatureAIXLocalExecTLS :
                    "Produce a TOC-free local-exec TLS sequence for this function "
                    "for 64-bit AIX">;
 
+// Specifies that local-dynamic TLS accesses in any function with this target
+// attribute should use the optimized sequence (where the offset is an immediate
+// off the module-handle for which the linker might add fix-up code for if the
+// immediate is too large).
+def FeatureAIXLocalDynamicTLS :
+  SubtargetFeature<"aix-small-local-dynamic-tls", "HasAIXSmallLocalDynamicTLS",
+                   "true", "Produce a faster local-dynamic TLS sequence for this "
+                   "function for 64-bit AIX">;
+
 def FeaturePredictableSelectIsExpensive :
   SubtargetFeature<"predictable-select-expensive",
                    "PredictableSelectIsExpensive",

llvm/lib/Target/PowerPC/PPCAsmPrinter.cpp

@@ -803,7 +803,8 @@ void PPCAsmPrinter::emitInstruction(const MachineInstr *MI) {
   MCInst TmpInst;
   const bool IsPPC64 = Subtarget->isPPC64();
   const bool IsAIX = Subtarget->isAIXABI();
-  const bool HasAIXSmallLocalExecTLS = Subtarget->hasAIXSmallLocalExecTLS();
+  const bool HasAIXSmallLocalTLS = Subtarget->hasAIXSmallLocalExecTLS() ||
+                                   Subtarget->hasAIXSmallLocalDynamicTLS();
   const Module *M = MF->getFunction().getParent();
   PICLevel::Level PL = M->getPICLevel();
 
@@ -1612,19 +1613,19 @@ void PPCAsmPrinter::emitInstruction(const MachineInstr *MI) {
   case PPC::LFD:
   case PPC::STFD:
   case PPC::ADDI8: {
-    // A faster non-TOC-based local-exec sequence is represented by `addi`
-    // or a load/store instruction (that directly loads or stores off of the
-    // thread pointer) with an immediate operand having the MO_TPREL_FLAG.
+    // A faster non-TOC-based local-[exec|dynamic] sequence is represented by
+    // `addi` or a load/store instruction (that directly loads or stores off of
+    // the thread pointer) with an immediate operand having the MO_TPREL_FLAG.
     // Such instructions do not otherwise arise.
-    if (!HasAIXSmallLocalExecTLS)
+    if (!HasAIXSmallLocalTLS)
       break;
     bool IsMIADDI8 = MI->getOpcode() == PPC::ADDI8;
     unsigned OpNum = IsMIADDI8 ? 2 : 1;
     const MachineOperand &MO = MI->getOperand(OpNum);
     unsigned Flag = MO.getTargetFlags();
     if (Flag == PPCII::MO_TPREL_FLAG ||
         Flag == PPCII::MO_GOT_TPREL_PCREL_FLAG ||
-        Flag == PPCII::MO_TPREL_PCREL_FLAG) {
+        Flag == PPCII::MO_TPREL_PCREL_FLAG || Flag == PPCII::MO_TLSLD_FLAG) {
       LowerPPCMachineInstrToMCInst(MI, TmpInst, *this);
 
       const MCExpr *Expr = getAdjustedLocalExecExpr(MO, MO.getOffset());
@@ -1672,7 +1673,12 @@ const MCExpr *PPCAsmPrinter::getAdjustedLocalExecExpr(const MachineOperand &MO,
 
   assert(MO.isGlobal() && "Only expecting a global MachineOperand here!");
   const GlobalValue *GValue = MO.getGlobal();
-  assert(TM.getTLSModel(GValue) == TLSModel::LocalExec &&
+  // TODO: Handle the aix-small-local-dynamic-tls non-zero offset case.
+  TLSModel::Model Model = TM.getTLSModel(GValue);
+  if (Model == TLSModel::LocalDynamic) {
+    return nullptr;
+  }
+  assert(Model == TLSModel::LocalExec &&
          "Only local-exec accesses are handled!");
 
   bool IsGlobalADeclaration = GValue->isDeclarationForLinker();

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

@@ -153,10 +153,10 @@ static SDValue widenVec(SelectionDAG &DAG, SDValue Vec, const SDLoc &dl);
 
 static const char AIXSSPCanaryWordName[] = "__ssp_canary_word";
 
-// A faster local-exec TLS access sequence (enabled with the
-// -maix-small-local-exec-tls option) can be produced for TLS variables;
-// consistent with the IBM XL compiler, we apply a max size of slightly under
-// 32KB.
+// A faster local-[exec|dynamic] TLS access sequence (enabled with the
+// -maix-small-local-[exec|dynamic]-tls option) can be produced for TLS
+// variables; consistent with the IBM XL compiler, we apply a max size of
+// slightly under 32KB.
 constexpr uint64_t AIXSmallTlsPolicySizeLimit = 32751;
 
 // FIXME: Remove this once the bug has been fixed!
@@ -3434,6 +3434,14 @@ SDValue PPCTargetLowering::LowerGlobalTLSAddressAIX(SDValue Op,
   }
 
   if (Model == TLSModel::LocalDynamic) {
+    bool HasAIXSmallLocalDynamicTLS = Subtarget.hasAIXSmallLocalDynamicTLS();
+
+    // We do not implement the 32-bit version of the faster access sequence
+    // for local-dynamic that is controlled by -maix-small-local-dynamic-tls.
+    if (!Is64Bit && HasAIXSmallLocalDynamicTLS)
+      report_fatal_error("The small-local-dynamic TLS access sequence is "
+                         "currently only supported on AIX (64-bit mode).");
+
     // For local-dynamic on AIX, we need to generate one TOC entry for each
     // variable offset, and a single module-handle TOC entry for the entire
     // file.
@@ -3454,6 +3462,22 @@ SDValue PPCTargetLowering::LowerGlobalTLSAddressAIX(SDValue Op,
     SDValue ModuleHandle =
         DAG.getNode(PPCISD::TLSLD_AIX, dl, PtrVT, ModuleHandleTOC);
 
+    // With the -maix-small-local-dynamic-tls option, produce a faster access
+    // sequence for local-dynamic TLS variables where the offset from the
+    // module-handle is encoded as an immediate operand.
+    //
+    // We only utilize the faster local-dynamic access sequence when the TLS
+    // variable has a size within the policy limit. We treat types that are
+    // not sized or are empty as being over the policy size limit.
+    if (HasAIXSmallLocalDynamicTLS) {
+      Type *GVType = GV->getValueType();
+      if (GVType->isSized() && !GVType->isEmptyTy() &&
+          GV->getParent()->getDataLayout().getTypeAllocSize(GVType) <=
+              AIXSmallTlsPolicySizeLimit)
+        return DAG.getNode(PPCISD::Lo, dl, PtrVT, VariableOffsetTGA,
+                           ModuleHandle);
+    }
+
     return DAG.getNode(ISD::ADD, dl, PtrVT, ModuleHandle, VariableOffset);
   }
 

llvm/lib/Target/PowerPC/PPCMCInstLower.cpp

@@ -96,15 +96,18 @@ static MCOperand GetSymbolRef(const MachineOperand &MO, const MCSymbol *Symbol,
     RefKind = MCSymbolRefExpr::VK_PPC_GOT_TLSLD_PCREL;
   else if (MO.getTargetFlags() == PPCII::MO_GOT_TPREL_PCREL_FLAG)
     RefKind = MCSymbolRefExpr::VK_PPC_GOT_TPREL_PCREL;
-  else if (MO.getTargetFlags() == PPCII::MO_TPREL_FLAG) {
+  else if (MO.getTargetFlags() == PPCII::MO_TPREL_FLAG ||
+           MO.getTargetFlags() == PPCII::MO_TLSLD_FLAG) {
     assert(MO.isGlobal() && "Only expecting a global MachineOperand here!");
     TLSModel::Model Model = TM.getTLSModel(MO.getGlobal());
-    // For the local-exec TLS model, we may generate the offset from the TLS
-    // base as an immediate operand (instead of using a TOC entry).
-    // Set the relocation type in case the result is used for purposes other
-    // than a TOC reference. In TOC reference cases, this result is discarded.
+    // For the local-[exec|dynamic] TLS model, we may generate the offset from
+    // the TLS base as an immediate operand (instead of using a TOC entry). Set
+    // the relocation type in case the result is used for purposes other than a
+    // TOC reference. In TOC reference cases, this result is discarded.
     if (Model == TLSModel::LocalExec)
       RefKind = MCSymbolRefExpr::VK_PPC_AIX_TLSLE;
+    else if (Model == TLSModel::LocalDynamic)
+      RefKind = MCSymbolRefExpr::VK_PPC_AIX_TLSLD;
   }
 
   const MachineInstr *MI = MO.getParent();

llvm/lib/Target/PowerPC/PPCSubtarget.cpp

@@ -124,22 +124,22 @@ void PPCSubtarget::initSubtargetFeatures(StringRef CPU, StringRef TuneCPU,
   // Determine endianness.
   IsLittleEndian = TM.isLittleEndian();
 
-  if (HasAIXSmallLocalExecTLS) {
+  if (HasAIXSmallLocalExecTLS || HasAIXSmallLocalDynamicTLS) {
     if (!TargetTriple.isOSAIX() || !IsPPC64)
-      report_fatal_error(
-          "The aix-small-local-exec-tls attribute is only supported on AIX in "
-          "64-bit mode.\n",
-          false);
-    // The aix-small-local-exec-tls attribute should only be used with
+      report_fatal_error("The aix-small-local-[exec|dynamic]-tls attribute is "
+                         "only supported on AIX in "
+                         "64-bit mode.\n",
+                         false);
+    // The aix-small-local-[exec|dynamic]-tls attribute should only be used with
     // -data-sections, as having data sections turned off with this option
-    // is not ideal for performance. Moreover, the small-local-exec-tls region
-    // is a limited resource, and should not be used for variables that may
-    // be replaced.
+    // is not ideal for performance. Moreover, the
+    // small-local-[exec|dynamic]-tls region is a limited resource, and should
+    // not be used for variables that may be replaced.
     if (!TM.getDataSections())
-      report_fatal_error(
-          "The aix-small-local-exec-tls attribute can only be specified with "
-          "-data-sections.\n",
-          false);
+      report_fatal_error("The aix-small-local-[exec|dynamic]-tls attribute can "
+                         "only be specified with "
+                         "-data-sections.\n",
+                         false);
   }
 }