diff --git a/src/mono/mono/mini/intrinsics.c b/src/mono/mono/mini/intrinsics.c
index 7e026e0e68bd98..17a9c14f81b464 100644
--- a/src/mono/mono/mini/intrinsics.c
+++ b/src/mono/mono/mini/intrinsics.c
@@ -297,8 +297,8 @@ llvm_emit_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSign
 		//
 		// * MinNumber / MaxNumber: IEEE 754-2019 minimumNumber / maximumNumber
 		//   (NaN-suppressing and sign-of-zero aware, treating -0 as less than +0).
-		//   Lowered in mini-llvm.c by composing llvm.minimum / llvm.maximum with an
-		//   explicit NaN fixup; see the OP_FMINNUM case there and llvm-intrinsics.h.
+		//   Lower to llvm.minimumnum / llvm.maximumnum, which on AArch64 maps to a
+		//   single fminnm / fmaxnm instruction; see mini-llvm.c and llvm-intrinsics.h.
 		// * Abs: BCL forwarder to MathF.Abs / Math.Abs. Today this usually inlines
 		//   into the Math/MathF recognition above, but adding direct recognition
 		//   keeps the lowering working even if the JIT inliner declines.
diff --git a/src/mono/mono/mini/llvm-intrinsics.h b/src/mono/mono/mini/llvm-intrinsics.h
index 848ef32a64bf82..ce09bacbd80b4a 100644
--- a/src/mono/mono/mini/llvm-intrinsics.h
+++ b/src/mono/mono/mini/llvm-intrinsics.h
@@ -98,11 +98,14 @@ INTRINS_OVR(TRUNCF, trunc, Generic, LLVMFloatType ())
 INTRINS_OVR(COPYSIGN, copysign, Generic, LLVMDoubleType ())
 INTRINS_OVR(COPYSIGNF, copysign, Generic, LLVMFloatType ())
 	/*
-	 * `float.MinNumber` / `double.MinNumber` (and the Max variants) are lowered in
-	 * mini-llvm.c by composing llvm.minimum/maximum (above) with an explicit NaN
-	 * fixup; see the OP_FMINNUM case there for why we don't use llvm.minnum/maxnum
-	 * or llvm.minimumnum/maximumnum directly.
+	 * IEEE 754-2019 minimumNumber/maximumNumber (NaN-suppressing and sign-of-zero
+	 * aware). Use llvm.minimum/maximum (above) for the NaN-propagating
+	 * Math.Min/Math.Max instead.
 	 */
+INTRINS_OVR(MINIMUMNUM, minimumnum, Generic, LLVMDoubleType ())
+INTRINS_OVR(MINIMUMNUMF, minimumnum, Generic, LLVMFloatType ())
+INTRINS_OVR(MAXIMUMNUM, maximumnum, Generic, LLVMDoubleType ())
+INTRINS_OVR(MAXIMUMNUMF, maximumnum, Generic, LLVMFloatType ())
 INTRINS_OVR(EXPECT_I8, expect, Generic, LLVMInt8Type ())
 INTRINS_OVR(EXPECT_I1, expect, Generic, LLVMInt1Type ())
 INTRINS_OVR(CTPOP_I32, ctpop, Generic, LLVMInt32Type ())
diff --git a/src/mono/mono/mini/mini-llvm.c b/src/mono/mono/mini/mini-llvm.c
index f3166c3d2881f1..7f97befdc31241 100644
--- a/src/mono/mono/mini/mini-llvm.c
+++ b/src/mono/mono/mini/mini-llvm.c
@@ -7628,32 +7628,20 @@ MONO_RESTORE_WARNING
 		case OP_RMAXNUM: {
 			/*
 			 * IEEE 754-2019 minimumNumber/maximumNumber (NaN-suppressing and
-			 * sign-of-zero aware), as specified by `float.MinNumber` /
-			 * `double.MinNumber` (and the Max variants, surfaced via
-			 * INumber<TSelf> on the primitive Single/Double/Half types).
-			 *
-			 * We compose this from llvm.minimum/maximum (sign-of-zero aware but
-			 * NaN-propagating) plus an explicit NaN fixup, rather than lowering
-			 * directly to an intrinsic: llvm.minnum/maxnum leave the sign of zero
-			 * unspecified (minnum(+0, -0) may return +0 on x86, see dotnet/runtime
-			 * #131130) and llvm.minimumnum/maximumnum are miscompiled by the x86
-			 * backend in the LLVM version we build against. When exactly one operand
-			 * is NaN we return the other; when both are NaN the NaN flows through.
+			 * sign-of-zero aware). Maps directly to llvm.minimumnum/maximumnum.
 			 */
 			gboolean is_r4 = ins->opcode == OP_RMINNUM || ins->opcode == OP_RMAXNUM;
-			gboolean is_max = ins->opcode == OP_FMAXNUM || ins->opcode == OP_RMAXNUM;
 			LLVMTypeRef t = is_r4 ? LLVMFloatType () : LLVMDoubleType ();
-			LLVMValueRef l = convert (ctx, lhs, t);
-			LLVMValueRef r = convert (ctx, rhs, t);
-			LLVMValueRef args [2] = { l, r };
-			IntrinsicId iid = is_max ? (is_r4 ? INTRINS_MAXIMUMF : INTRINS_MAXIMUM)
-			                         : (is_r4 ? INTRINS_MINIMUMF : INTRINS_MINIMUM);
-			LLVMValueRef result = call_intrins (ctx, iid, args, "");
-			LLVMValueRef l_nan = LLVMBuildFCmp (builder, LLVMRealUNO, l, l, "");
-			LLVMValueRef r_nan = LLVMBuildFCmp (builder, LLVMRealUNO, r, r, "");
-			result = LLVMBuildSelect (builder, r_nan, l, result, "");
-			result = LLVMBuildSelect (builder, l_nan, r, result, dname);
-			values [ins->dreg] = result;
+			LLVMValueRef args [2] = { convert (ctx, lhs, t), convert (ctx, rhs, t) };
+			IntrinsicId iid;
+			switch (ins->opcode) {
+			case OP_FMAXNUM: iid = INTRINS_MAXIMUMNUM; break;
+			case OP_FMINNUM: iid = INTRINS_MINIMUMNUM; break;
+			case OP_RMAXNUM: iid = INTRINS_MAXIMUMNUMF; break;
+			case OP_RMINNUM: iid = INTRINS_MINIMUMNUMF; break;
+			default: g_assert_not_reached (); break;
+			}
+			values [ins->dreg] = call_intrins (ctx, iid, args, dname);
 			break;
 		}