AArch64: Use consistent atomicrmw expansion for FP operations (#103702)

Use LLSC or cmpxchg in the same cases as for the unsupported
integer operations. This required some fixups to the LLSC
implementatation to deal with the fp128 case.

The comment about floating-point exceptions was wrong,
because floating-point exceptions are not really exceptions at all.

Author: arsenm

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -27096,21 +27096,37 @@ AArch64TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
                              : AtomicExpansionKind::LLSC;
 }
 
+// Return true if the atomic operation expansion will lower to use a library
+// call, and is thus ineligible to use an LLSC expansion.
+static bool rmwOpMayLowerToLibcall(const AArch64Subtarget &Subtarget,
+                                   const AtomicRMWInst *RMW) {
+  if (!RMW->isFloatingPointOperation())
+    return false;
+  switch (RMW->getType()->getScalarType()->getTypeID()) {
+  case Type::FloatTyID:
+  case Type::DoubleTyID:
+  case Type::HalfTyID:
+  case Type::BFloatTyID:
+    // Will use soft float
+    return !Subtarget.hasFPARMv8();
+  default:
+    // fp128 will emit library calls.
+    return true;
+  }
+
+  llvm_unreachable("covered type switch");
+}
+
 // The "default" for integer RMW operations is to expand to an LL/SC loop.
 // However, with the LSE instructions (or outline-atomics mode, which provides
 // library routines in place of the LSE-instructions), we can directly emit many
 // operations instead.
-//
-// Floating-point operations are always emitted to a cmpxchg loop, because they
-// may trigger a trap which aborts an LLSC sequence.
 TargetLowering::AtomicExpansionKind
 AArch64TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
-  unsigned Size = AI->getType()->getPrimitiveSizeInBits();
+  Type *Ty = AI->getType();
+  unsigned Size = Ty->getPrimitiveSizeInBits();
   assert(Size <= 128 && "AtomicExpandPass should've handled larger sizes.");
 
-  if (AI->isFloatingPointOperation())
-    return AtomicExpansionKind::CmpXChg;
-
   bool CanUseLSE128 = Subtarget->hasLSE128() && Size == 128 &&
                       (AI->getOperation() == AtomicRMWInst::Xchg ||
                        AI->getOperation() == AtomicRMWInst::Or ||
@@ -27120,7 +27136,8 @@ AArch64TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
 
   // Nand is not supported in LSE.
   // Leave 128 bits to LLSC or CmpXChg.
-  if (AI->getOperation() != AtomicRMWInst::Nand && Size < 128) {
+  if (AI->getOperation() != AtomicRMWInst::Nand && Size < 128 &&
+      !AI->isFloatingPointOperation()) {
     if (Subtarget->hasLSE())
       return AtomicExpansionKind::None;
     if (Subtarget->outlineAtomics()) {
@@ -27146,7 +27163,7 @@ AArch64TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
   // succeed. So at -O0 lower this operation to a CAS loop. Also worthwhile if
   // we have a single CAS instruction that can replace the loop.
   if (getTargetMachine().getOptLevel() == CodeGenOptLevel::None ||
-      Subtarget->hasLSE())
+      Subtarget->hasLSE() || rmwOpMayLowerToLibcall(*Subtarget, AI))
     return AtomicExpansionKind::CmpXChg;
 
   return AtomicExpansionKind::LLSC;
@@ -27193,10 +27210,14 @@ Value *AArch64TargetLowering::emitLoadLinked(IRBuilderBase &Builder,
 
     Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
     Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
-    Lo = Builder.CreateZExt(Lo, ValueTy, "lo64");
-    Hi = Builder.CreateZExt(Hi, ValueTy, "hi64");
-    return Builder.CreateOr(
-        Lo, Builder.CreateShl(Hi, ConstantInt::get(ValueTy, 64)), "val64");
+
+    auto *Int128Ty = Type::getInt128Ty(Builder.getContext());
+    Lo = Builder.CreateZExt(Lo, Int128Ty, "lo64");
+    Hi = Builder.CreateZExt(Hi, Int128Ty, "hi64");
+
+    Value *Or = Builder.CreateOr(
+        Lo, Builder.CreateShl(Hi, ConstantInt::get(Int128Ty, 64)), "val64");
+    return Builder.CreateBitCast(Or, ValueTy);
   }
 
   Type *Tys[] = { Addr->getType() };
@@ -27207,8 +27228,8 @@ Value *AArch64TargetLowering::emitLoadLinked(IRBuilderBase &Builder,
   const DataLayout &DL = M->getDataLayout();
   IntegerType *IntEltTy = Builder.getIntNTy(DL.getTypeSizeInBits(ValueTy));
   CallInst *CI = Builder.CreateCall(Ldxr, Addr);
-  CI->addParamAttr(
-      0, Attribute::get(Builder.getContext(), Attribute::ElementType, ValueTy));
+  CI->addParamAttr(0, Attribute::get(Builder.getContext(),
+                                     Attribute::ElementType, IntEltTy));
   Value *Trunc = Builder.CreateTrunc(CI, IntEltTy);
 
   return Builder.CreateBitCast(Trunc, ValueTy);
@@ -27234,9 +27255,13 @@ Value *AArch64TargetLowering::emitStoreConditional(IRBuilderBase &Builder,
         IsRelease ? Intrinsic::aarch64_stlxp : Intrinsic::aarch64_stxp;
     Function *Stxr = Intrinsic::getDeclaration(M, Int);
     Type *Int64Ty = Type::getInt64Ty(M->getContext());
+    Type *Int128Ty = Type::getInt128Ty(M->getContext());
+
+    Value *CastVal = Builder.CreateBitCast(Val, Int128Ty);
 
-    Value *Lo = Builder.CreateTrunc(Val, Int64Ty, "lo");
-    Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 64), Int64Ty, "hi");
+    Value *Lo = Builder.CreateTrunc(CastVal, Int64Ty, "lo");
+    Value *Hi =
+        Builder.CreateTrunc(Builder.CreateLShr(CastVal, 64), Int64Ty, "hi");
     return Builder.CreateCall(Stxr, {Lo, Hi, Addr});
   }