[AArch64][llvm] Codegen for 16/32/64-bit floating-point atomicrmw fminimum/faximum (#137703)

Codegen for AArch64 16/32/64-bit floating-point atomic read-modify-write
operations (`atomicrmw {fmaximum,fminimum}`) using LD{B}FMAX and
LD{B}FMIN atomic instructions.

Author: jthackray

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -1001,6 +1001,16 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
     setOperationAction(ISD::ATOMIC_LOAD_FMIN, MVT::f32, LibCall);
     setOperationAction(ISD::ATOMIC_LOAD_FMIN, MVT::f64, LibCall);
     setOperationAction(ISD::ATOMIC_LOAD_FMIN, MVT::bf16, LibCall);
+
+    setOperationAction(ISD::ATOMIC_LOAD_FMAXIMUM, MVT::f16, LibCall);
+    setOperationAction(ISD::ATOMIC_LOAD_FMAXIMUM, MVT::f32, LibCall);
+    setOperationAction(ISD::ATOMIC_LOAD_FMAXIMUM, MVT::f64, LibCall);
+    setOperationAction(ISD::ATOMIC_LOAD_FMAXIMUM, MVT::bf16, LibCall);
+
+    setOperationAction(ISD::ATOMIC_LOAD_FMINIMUM, MVT::f16, LibCall);
+    setOperationAction(ISD::ATOMIC_LOAD_FMINIMUM, MVT::f32, LibCall);
+    setOperationAction(ISD::ATOMIC_LOAD_FMINIMUM, MVT::f64, LibCall);
+    setOperationAction(ISD::ATOMIC_LOAD_FMINIMUM, MVT::bf16, LibCall);
   }
 
   if (Subtarget->hasLSE128()) {
@@ -27991,7 +28001,9 @@ AArch64TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
   // If LSFE available, use atomic FP instructions in preference to expansion
   if (Subtarget->hasLSFE() && (AI->getOperation() == AtomicRMWInst::FAdd ||
                                AI->getOperation() == AtomicRMWInst::FMax ||
-                               AI->getOperation() == AtomicRMWInst::FMin))
+                               AI->getOperation() == AtomicRMWInst::FMin ||
+                               AI->getOperation() == AtomicRMWInst::FMaximum ||
+                               AI->getOperation() == AtomicRMWInst::FMinimum))
     return AtomicExpansionKind::None;
 
   // Nand is not supported in LSE.

llvm/lib/Target/AArch64/AArch64InstrAtomics.td

@@ -551,14 +551,21 @@ defm atomic_load_fadd  : binary_atomic_op_fp<atomic_load_fadd>;
 defm atomic_load_fmin  : binary_atomic_op_fp<atomic_load_fmin>;
 defm atomic_load_fmax  : binary_atomic_op_fp<atomic_load_fmax>;
 
+defm atomic_load_fminimum  : binary_atomic_op_fp<atomic_load_fminimum>;
+defm atomic_load_fmaximum  : binary_atomic_op_fp<atomic_load_fmaximum>;
+
 let Predicates = [HasLSFE] in {
   defm : LDFPOPregister_patterns<"LDFADD",   "atomic_load_fadd">;
   defm : LDFPOPregister_patterns<"LDFMAXNM", "atomic_load_fmax">;
   defm : LDFPOPregister_patterns<"LDFMINNM", "atomic_load_fmin">;
+  defm : LDFPOPregister_patterns<"LDFMAX",   "atomic_load_fmaximum">;
+  defm : LDFPOPregister_patterns<"LDFMIN",   "atomic_load_fminimum">;
 
   defm : LDBFPOPregister_patterns<"LDBFADD",   "atomic_load_fadd">;
   defm : LDBFPOPregister_patterns<"LDBFMAXNM", "atomic_load_fmax">;
   defm : LDBFPOPregister_patterns<"LDBFMINNM", "atomic_load_fmin">;
+  defm : LDBFPOPregister_patterns<"LDBFMAX",   "atomic_load_fmaximum">;
+  defm : LDBFPOPregister_patterns<"LDBFMIN",   "atomic_load_fminimum">;
 }
 
 // v8.9a/v9.4a FEAT_LRCPC patterns