; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py ; RUN: llc < %s -mtriple aarch64 | FileCheck %s --check-prefixes=CHECK-LE ; RUN: llc < %s -mtriple aarch64_be | FileCheck %s --check-prefixes=CHECK-BE ; Note: These tests show that LDNP is not (yet) prioritized for interleaved loads. define void @test_ldnp_interleaved_load2_v2i32(ptr %ptr, ptr %out0, ptr %out1) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load2_v2i32: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld2 { v0.2s, v1.2s }, [x0] ; CHECK-LE-NEXT: str d0, [x1] ; CHECK-LE-NEXT: str d1, [x2] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load2_v2i32: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld2 { v0.2s, v1.2s }, [x0] ; CHECK-BE-NEXT: st1 { v0.2s }, [x1] ; CHECK-BE-NEXT: st1 { v1.2s }, [x2] ; CHECK-BE-NEXT: ret entry: %loaded = load <4 x i32>, ptr %ptr, align 4, !nontemporal !0 %v0 = shufflevector <4 x i32> %loaded, <4 x i32> poison, <2 x i32> %v1 = shufflevector <4 x i32> %loaded, <4 x i32> poison, <2 x i32> store <2 x i32> %v0, ptr %out0 store <2 x i32> %v1, ptr %out1 ret void } define void @test_ldnp_interleaved_load2_v4i16(ptr %ptr, ptr %out0, ptr %out1) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load2_v4i16: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld2 { v0.4h, v1.4h }, [x0] ; CHECK-LE-NEXT: str d0, [x1] ; CHECK-LE-NEXT: str d1, [x2] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load2_v4i16: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld2 { v0.4h, v1.4h }, [x0] ; CHECK-BE-NEXT: st1 { v0.4h }, [x1] ; CHECK-BE-NEXT: st1 { v1.4h }, [x2] ; CHECK-BE-NEXT: ret entry: %loaded = load <8 x i16>, ptr %ptr, align 2, !nontemporal !0 %v0 = shufflevector <8 x i16> %loaded, <8 x i16> poison, <4 x i32> %v1 = shufflevector <8 x i16> %loaded, <8 x i16> poison, <4 x i32> store <4 x i16> %v0, ptr %out0 store <4 x i16> %v1, ptr %out1 ret void } define void @test_ldnp_interleaved_load2_v8i8(ptr %ptr, ptr %out0, ptr %out1) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load2_v8i8: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld2 { v0.8b, v1.8b }, [x0] ; CHECK-LE-NEXT: str d0, [x1] ; CHECK-LE-NEXT: str d1, [x2] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load2_v8i8: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld2 { v0.8b, v1.8b }, [x0] ; CHECK-BE-NEXT: st1 { v0.8b }, [x1] ; CHECK-BE-NEXT: st1 { v1.8b }, [x2] ; CHECK-BE-NEXT: ret entry: %loaded = load <16 x i8>, ptr %ptr, align 1, !nontemporal !0 %v0 = shufflevector <16 x i8> %loaded, <16 x i8> poison, <8 x i32> %v1 = shufflevector <16 x i8> %loaded, <16 x i8> poison, <8 x i32> store <8 x i8> %v0, ptr %out0 store <8 x i8> %v1, ptr %out1 ret void } define void @test_ldnp_interleaved_load2_v2f32(ptr %ptr, ptr %out0, ptr %out1) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load2_v2f32: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld2 { v0.2s, v1.2s }, [x0] ; CHECK-LE-NEXT: str d0, [x1] ; CHECK-LE-NEXT: str d1, [x2] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load2_v2f32: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld2 { v0.2s, v1.2s }, [x0] ; CHECK-BE-NEXT: st1 { v0.2s }, [x1] ; CHECK-BE-NEXT: st1 { v1.2s }, [x2] ; CHECK-BE-NEXT: ret entry: %loaded = load <4 x float>, ptr %ptr, align 4, !nontemporal !0 %v0 = shufflevector <4 x float> %loaded, <4 x float> poison, <2 x i32> %v1 = shufflevector <4 x float> %loaded, <4 x float> poison, <2 x i32> store <2 x float> %v0, ptr %out0 store <2 x float> %v1, ptr %out1 ret void } define void @test_ldnp_interleaved_load2_v4f16(ptr %ptr, ptr %out0, ptr %out1) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load2_v4f16: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld2 { v0.4h, v1.4h }, [x0] ; CHECK-LE-NEXT: str d0, [x1] ; CHECK-LE-NEXT: str d1, [x2] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load2_v4f16: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld2 { v0.4h, v1.4h }, [x0] ; CHECK-BE-NEXT: st1 { v0.4h }, [x1] ; CHECK-BE-NEXT: st1 { v1.4h }, [x2] ; CHECK-BE-NEXT: ret entry: %loaded = load <8 x half>, ptr %ptr, align 2, !nontemporal !0 %v0 = shufflevector <8 x half> %loaded, <8 x half> poison, <4 x i32> %v1 = shufflevector <8 x half> %loaded, <8 x half> poison, <4 x i32> store <4 x half> %v0, ptr %out0 store <4 x half> %v1, ptr %out1 ret void } define void @test_ldnp_interleaved_load2_v2i64(ptr %ptr, ptr %out0, ptr %out1) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load2_v2i64: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld2 { v0.2d, v1.2d }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load2_v2i64: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld2 { v0.2d, v1.2d }, [x0] ; CHECK-BE-NEXT: st1 { v0.2d }, [x1] ; CHECK-BE-NEXT: st1 { v1.2d }, [x2] ; CHECK-BE-NEXT: ret entry: %loaded = load <4 x i64>, ptr %ptr, align 8, !nontemporal !0 %v0 = shufflevector <4 x i64> %loaded, <4 x i64> poison, <2 x i32> %v1 = shufflevector <4 x i64> %loaded, <4 x i64> poison, <2 x i32> store <2 x i64> %v0, ptr %out0 store <2 x i64> %v1, ptr %out1 ret void } define void @test_ldnp_interleaved_load2_v4i32(ptr %ptr, ptr %out0, ptr %out1) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load2_v4i32: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld2 { v0.4s, v1.4s }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load2_v4i32: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld2 { v0.4s, v1.4s }, [x0] ; CHECK-BE-NEXT: st1 { v0.4s }, [x1] ; CHECK-BE-NEXT: st1 { v1.4s }, [x2] ; CHECK-BE-NEXT: ret entry: %loaded = load <8 x i32>, ptr %ptr, align 4, !nontemporal !0 %v0 = shufflevector <8 x i32> %loaded, <8 x i32> poison, <4 x i32> %v1 = shufflevector <8 x i32> %loaded, <8 x i32> poison, <4 x i32> store <4 x i32> %v0, ptr %out0 store <4 x i32> %v1, ptr %out1 ret void } define void @test_ldnp_interleaved_load2_v8i16(ptr %ptr, ptr %out0, ptr %out1) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load2_v8i16: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld2 { v0.8h, v1.8h }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load2_v8i16: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld2 { v0.8h, v1.8h }, [x0] ; CHECK-BE-NEXT: st1 { v0.8h }, [x1] ; CHECK-BE-NEXT: st1 { v1.8h }, [x2] ; CHECK-BE-NEXT: ret entry: %loaded = load <16 x i16>, ptr %ptr, align 2, !nontemporal !0 %v0 = shufflevector <16 x i16> %loaded, <16 x i16> poison, <8 x i32> %v1 = shufflevector <16 x i16> %loaded, <16 x i16> poison, <8 x i32> store <8 x i16> %v0, ptr %out0 store <8 x i16> %v1, ptr %out1 ret void } define void @test_ldnp_interleaved_load2_v16i8(ptr %ptr, ptr %out0, ptr %out1) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load2_v16i8: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld2 { v0.16b, v1.16b }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load2_v16i8: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld2 { v0.16b, v1.16b }, [x0] ; CHECK-BE-NEXT: st1 { v0.16b }, [x1] ; CHECK-BE-NEXT: st1 { v1.16b }, [x2] ; CHECK-BE-NEXT: ret entry: %loaded = load <32 x i8>, ptr %ptr, align 1, !nontemporal !0 %v0 = shufflevector <32 x i8> %loaded, <32 x i8> poison, <16 x i32> %v1 = shufflevector <32 x i8> %loaded, <32 x i8> poison, <16 x i32> store <16 x i8> %v0, ptr %out0 store <16 x i8> %v1, ptr %out1 ret void } define void @test_ldnp_interleaved_load2_v2f64(ptr %ptr, ptr %out0, ptr %out1) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load2_v2f64: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld2 { v0.2d, v1.2d }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load2_v2f64: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld2 { v0.2d, v1.2d }, [x0] ; CHECK-BE-NEXT: st1 { v0.2d }, [x1] ; CHECK-BE-NEXT: st1 { v1.2d }, [x2] ; CHECK-BE-NEXT: ret entry: %loaded = load <4 x double>, ptr %ptr, align 8, !nontemporal !0 %v0 = shufflevector <4 x double> %loaded, <4 x double> poison, <2 x i32> %v1 = shufflevector <4 x double> %loaded, <4 x double> poison, <2 x i32> store <2 x double> %v0, ptr %out0 store <2 x double> %v1, ptr %out1 ret void } define void @test_ldnp_interleaved_load2_v4f32(ptr %ptr, ptr %out0, ptr %out1) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load2_v4f32: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld2 { v0.4s, v1.4s }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load2_v4f32: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld2 { v0.4s, v1.4s }, [x0] ; CHECK-BE-NEXT: st1 { v0.4s }, [x1] ; CHECK-BE-NEXT: st1 { v1.4s }, [x2] ; CHECK-BE-NEXT: ret entry: %loaded = load <8 x float>, ptr %ptr, align 4, !nontemporal !0 %v0 = shufflevector <8 x float> %loaded, <8 x float> poison, <4 x i32> %v1 = shufflevector <8 x float> %loaded, <8 x float> poison, <4 x i32> store <4 x float> %v0, ptr %out0 store <4 x float> %v1, ptr %out1 ret void } define void @test_ldnp_interleaved_load2_v8f16(ptr %ptr, ptr %out0, ptr %out1) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load2_v8f16: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld2 { v0.8h, v1.8h }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load2_v8f16: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld2 { v0.8h, v1.8h }, [x0] ; CHECK-BE-NEXT: st1 { v0.8h }, [x1] ; CHECK-BE-NEXT: st1 { v1.8h }, [x2] ; CHECK-BE-NEXT: ret entry: %loaded = load <16 x half>, ptr %ptr, align 2, !nontemporal !0 %v0 = shufflevector <16 x half> %loaded, <16 x half> poison, <8 x i32> %v1 = shufflevector <16 x half> %loaded, <16 x half> poison, <8 x i32> store <8 x half> %v0, ptr %out0 store <8 x half> %v1, ptr %out1 ret void } define void @test_ldnp_interleaved_load3_v2i32(ptr %ptr, ptr %out0, ptr %out1, ptr %out2) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load3_v2i32: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld3 { v0.2s, v1.2s, v2.2s }, [x0] ; CHECK-LE-NEXT: str d0, [x1] ; CHECK-LE-NEXT: str d1, [x2] ; CHECK-LE-NEXT: str d2, [x3] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load3_v2i32: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld3 { v0.2s, v1.2s, v2.2s }, [x0] ; CHECK-BE-NEXT: st1 { v0.2s }, [x1] ; CHECK-BE-NEXT: st1 { v1.2s }, [x2] ; CHECK-BE-NEXT: st1 { v2.2s }, [x3] ; CHECK-BE-NEXT: ret entry: %loaded = load <6 x i32>, ptr %ptr, align 4, !nontemporal !0 %v0 = shufflevector <6 x i32> %loaded, <6 x i32> poison, <2 x i32> %v1 = shufflevector <6 x i32> %loaded, <6 x i32> poison, <2 x i32> %v2 = shufflevector <6 x i32> %loaded, <6 x i32> poison, <2 x i32> store <2 x i32> %v0, ptr %out0 store <2 x i32> %v1, ptr %out1 store <2 x i32> %v2, ptr %out2 ret void } define void @test_ldnp_interleaved_load3_v4i16(ptr %ptr, ptr %out0, ptr %out1, ptr %out2) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load3_v4i16: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld3 { v0.4h, v1.4h, v2.4h }, [x0] ; CHECK-LE-NEXT: str d0, [x1] ; CHECK-LE-NEXT: str d1, [x2] ; CHECK-LE-NEXT: str d2, [x3] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load3_v4i16: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld3 { v0.4h, v1.4h, v2.4h }, [x0] ; CHECK-BE-NEXT: st1 { v0.4h }, [x1] ; CHECK-BE-NEXT: st1 { v1.4h }, [x2] ; CHECK-BE-NEXT: st1 { v2.4h }, [x3] ; CHECK-BE-NEXT: ret entry: %loaded = load <12 x i16>, ptr %ptr, align 2, !nontemporal !0 %v0 = shufflevector <12 x i16> %loaded, <12 x i16> poison, <4 x i32> %v1 = shufflevector <12 x i16> %loaded, <12 x i16> poison, <4 x i32> %v2 = shufflevector <12 x i16> %loaded, <12 x i16> poison, <4 x i32> store <4 x i16> %v0, ptr %out0 store <4 x i16> %v1, ptr %out1 store <4 x i16> %v2, ptr %out2 ret void } define void @test_ldnp_interleaved_load3_v8i8(ptr %ptr, ptr %out0, ptr %out1, ptr %out2) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load3_v8i8: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld3 { v0.8b, v1.8b, v2.8b }, [x0] ; CHECK-LE-NEXT: str d0, [x1] ; CHECK-LE-NEXT: str d1, [x2] ; CHECK-LE-NEXT: str d2, [x3] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load3_v8i8: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld3 { v0.8b, v1.8b, v2.8b }, [x0] ; CHECK-BE-NEXT: st1 { v0.8b }, [x1] ; CHECK-BE-NEXT: st1 { v1.8b }, [x2] ; CHECK-BE-NEXT: st1 { v2.8b }, [x3] ; CHECK-BE-NEXT: ret entry: %loaded = load <24 x i8>, ptr %ptr, align 1, !nontemporal !0 %v0 = shufflevector <24 x i8> %loaded, <24 x i8> poison, <8 x i32> %v1 = shufflevector <24 x i8> %loaded, <24 x i8> poison, <8 x i32> %v2 = shufflevector <24 x i8> %loaded, <24 x i8> poison, <8 x i32> store <8 x i8> %v0, ptr %out0 store <8 x i8> %v1, ptr %out1 store <8 x i8> %v2, ptr %out2 ret void } define void @test_ldnp_interleaved_load3_v2f32(ptr %ptr, ptr %out0, ptr %out1, ptr %out2) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load3_v2f32: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld3 { v0.2s, v1.2s, v2.2s }, [x0] ; CHECK-LE-NEXT: str d0, [x1] ; CHECK-LE-NEXT: str d1, [x2] ; CHECK-LE-NEXT: str d2, [x3] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load3_v2f32: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld3 { v0.2s, v1.2s, v2.2s }, [x0] ; CHECK-BE-NEXT: st1 { v0.2s }, [x1] ; CHECK-BE-NEXT: st1 { v1.2s }, [x2] ; CHECK-BE-NEXT: st1 { v2.2s }, [x3] ; CHECK-BE-NEXT: ret entry: %loaded = load <6 x float>, ptr %ptr, align 4, !nontemporal !0 %v0 = shufflevector <6 x float> %loaded, <6 x float> poison, <2 x i32> %v1 = shufflevector <6 x float> %loaded, <6 x float> poison, <2 x i32> %v2 = shufflevector <6 x float> %loaded, <6 x float> poison, <2 x i32> store <2 x float> %v0, ptr %out0 store <2 x float> %v1, ptr %out1 store <2 x float> %v2, ptr %out2 ret void } define void @test_ldnp_interleaved_load3_v4f16(ptr %ptr, ptr %out0, ptr %out1, ptr %out2) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load3_v4f16: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld3 { v0.4h, v1.4h, v2.4h }, [x0] ; CHECK-LE-NEXT: str d0, [x1] ; CHECK-LE-NEXT: str d1, [x2] ; CHECK-LE-NEXT: str d2, [x3] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load3_v4f16: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld3 { v0.4h, v1.4h, v2.4h }, [x0] ; CHECK-BE-NEXT: st1 { v0.4h }, [x1] ; CHECK-BE-NEXT: st1 { v1.4h }, [x2] ; CHECK-BE-NEXT: st1 { v2.4h }, [x3] ; CHECK-BE-NEXT: ret entry: %loaded = load <12 x half>, ptr %ptr, align 2, !nontemporal !0 %v0 = shufflevector <12 x half> %loaded, <12 x half> poison, <4 x i32> %v1 = shufflevector <12 x half> %loaded, <12 x half> poison, <4 x i32> %v2 = shufflevector <12 x half> %loaded, <12 x half> poison, <4 x i32> store <4 x half> %v0, ptr %out0 store <4 x half> %v1, ptr %out1 store <4 x half> %v2, ptr %out2 ret void } define void @test_ldnp_interleaved_load3_v2i64(ptr %ptr, ptr %out0, ptr %out1, ptr %out2) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load3_v2i64: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld3 { v0.2d, v1.2d, v2.2d }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: str q2, [x3] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load3_v2i64: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld3 { v0.2d, v1.2d, v2.2d }, [x0] ; CHECK-BE-NEXT: st1 { v0.2d }, [x1] ; CHECK-BE-NEXT: st1 { v1.2d }, [x2] ; CHECK-BE-NEXT: st1 { v2.2d }, [x3] ; CHECK-BE-NEXT: ret entry: %loaded = load <6 x i64>, ptr %ptr, align 8, !nontemporal !0 %v0 = shufflevector <6 x i64> %loaded, <6 x i64> poison, <2 x i32> %v1 = shufflevector <6 x i64> %loaded, <6 x i64> poison, <2 x i32> %v2 = shufflevector <6 x i64> %loaded, <6 x i64> poison, <2 x i32> store <2 x i64> %v0, ptr %out0 store <2 x i64> %v1, ptr %out1 store <2 x i64> %v2, ptr %out2 ret void } define void @test_ldnp_interleaved_load3_v4i32(ptr %ptr, ptr %out0, ptr %out1, ptr %out2) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load3_v4i32: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld3 { v0.4s, v1.4s, v2.4s }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: str q2, [x3] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load3_v4i32: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld3 { v0.4s, v1.4s, v2.4s }, [x0] ; CHECK-BE-NEXT: st1 { v0.4s }, [x1] ; CHECK-BE-NEXT: st1 { v1.4s }, [x2] ; CHECK-BE-NEXT: st1 { v2.4s }, [x3] ; CHECK-BE-NEXT: ret entry: %loaded = load <12 x i32>, ptr %ptr, align 4, !nontemporal !0 %v0 = shufflevector <12 x i32> %loaded, <12 x i32> poison, <4 x i32> %v1 = shufflevector <12 x i32> %loaded, <12 x i32> poison, <4 x i32> %v2 = shufflevector <12 x i32> %loaded, <12 x i32> poison, <4 x i32> store <4 x i32> %v0, ptr %out0 store <4 x i32> %v1, ptr %out1 store <4 x i32> %v2, ptr %out2 ret void } define void @test_ldnp_interleaved_load3_v8i16(ptr %ptr, ptr %out0, ptr %out1, ptr %out2) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load3_v8i16: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld3 { v0.8h, v1.8h, v2.8h }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: str q2, [x3] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load3_v8i16: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld3 { v0.8h, v1.8h, v2.8h }, [x0] ; CHECK-BE-NEXT: st1 { v0.8h }, [x1] ; CHECK-BE-NEXT: st1 { v1.8h }, [x2] ; CHECK-BE-NEXT: st1 { v2.8h }, [x3] ; CHECK-BE-NEXT: ret entry: %loaded = load <24 x i16>, ptr %ptr, align 2, !nontemporal !0 %v0 = shufflevector <24 x i16> %loaded, <24 x i16> poison, <8 x i32> %v1 = shufflevector <24 x i16> %loaded, <24 x i16> poison, <8 x i32> %v2 = shufflevector <24 x i16> %loaded, <24 x i16> poison, <8 x i32> store <8 x i16> %v0, ptr %out0 store <8 x i16> %v1, ptr %out1 store <8 x i16> %v2, ptr %out2 ret void } define void @test_ldnp_interleaved_load3_v16i8(ptr %ptr, ptr %out0, ptr %out1, ptr %out2) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load3_v16i8: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld3 { v0.16b, v1.16b, v2.16b }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: str q2, [x3] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load3_v16i8: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld3 { v0.16b, v1.16b, v2.16b }, [x0] ; CHECK-BE-NEXT: st1 { v0.16b }, [x1] ; CHECK-BE-NEXT: st1 { v1.16b }, [x2] ; CHECK-BE-NEXT: st1 { v2.16b }, [x3] ; CHECK-BE-NEXT: ret entry: %loaded = load <48 x i8>, ptr %ptr, align 1, !nontemporal !0 %v0 = shufflevector <48 x i8> %loaded, <48 x i8> poison, <16 x i32> %v1 = shufflevector <48 x i8> %loaded, <48 x i8> poison, <16 x i32> %v2 = shufflevector <48 x i8> %loaded, <48 x i8> poison, <16 x i32> store <16 x i8> %v0, ptr %out0 store <16 x i8> %v1, ptr %out1 store <16 x i8> %v2, ptr %out2 ret void } define void @test_ldnp_interleaved_load3_v2f64(ptr %ptr, ptr %out0, ptr %out1, ptr %out2) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load3_v2f64: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld3 { v0.2d, v1.2d, v2.2d }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: str q2, [x3] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load3_v2f64: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld3 { v0.2d, v1.2d, v2.2d }, [x0] ; CHECK-BE-NEXT: st1 { v0.2d }, [x1] ; CHECK-BE-NEXT: st1 { v1.2d }, [x2] ; CHECK-BE-NEXT: st1 { v2.2d }, [x3] ; CHECK-BE-NEXT: ret entry: %loaded = load <6 x double>, ptr %ptr, align 8, !nontemporal !0 %v0 = shufflevector <6 x double> %loaded, <6 x double> poison, <2 x i32> %v1 = shufflevector <6 x double> %loaded, <6 x double> poison, <2 x i32> %v2 = shufflevector <6 x double> %loaded, <6 x double> poison, <2 x i32> store <2 x double> %v0, ptr %out0 store <2 x double> %v1, ptr %out1 store <2 x double> %v2, ptr %out2 ret void } define void @test_ldnp_interleaved_load3_v4f32(ptr %ptr, ptr %out0, ptr %out1, ptr %out2) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load3_v4f32: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld3 { v0.4s, v1.4s, v2.4s }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: str q2, [x3] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load3_v4f32: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld3 { v0.4s, v1.4s, v2.4s }, [x0] ; CHECK-BE-NEXT: st1 { v0.4s }, [x1] ; CHECK-BE-NEXT: st1 { v1.4s }, [x2] ; CHECK-BE-NEXT: st1 { v2.4s }, [x3] ; CHECK-BE-NEXT: ret entry: %loaded = load <12 x float>, ptr %ptr, align 4, !nontemporal !0 %v0 = shufflevector <12 x float> %loaded, <12 x float> poison, <4 x i32> %v1 = shufflevector <12 x float> %loaded, <12 x float> poison, <4 x i32> %v2 = shufflevector <12 x float> %loaded, <12 x float> poison, <4 x i32> store <4 x float> %v0, ptr %out0 store <4 x float> %v1, ptr %out1 store <4 x float> %v2, ptr %out2 ret void } define void @test_ldnp_interleaved_load3_v8f16(ptr %ptr, ptr %out0, ptr %out1, ptr %out2) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load3_v8f16: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld3 { v0.8h, v1.8h, v2.8h }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: str q2, [x3] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load3_v8f16: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld3 { v0.8h, v1.8h, v2.8h }, [x0] ; CHECK-BE-NEXT: st1 { v0.8h }, [x1] ; CHECK-BE-NEXT: st1 { v1.8h }, [x2] ; CHECK-BE-NEXT: st1 { v2.8h }, [x3] ; CHECK-BE-NEXT: ret entry: %loaded = load <24 x half>, ptr %ptr, align 2, !nontemporal !0 %v0 = shufflevector <24 x half> %loaded, <24 x half> poison, <8 x i32> %v1 = shufflevector <24 x half> %loaded, <24 x half> poison, <8 x i32> %v2 = shufflevector <24 x half> %loaded, <24 x half> poison, <8 x i32> store <8 x half> %v0, ptr %out0 store <8 x half> %v1, ptr %out1 store <8 x half> %v2, ptr %out2 ret void } ; Test conservative lowering of a ld4 matching patterns define void @test_ldnp_interleaved_load4_v2i32(ptr %ptr, ptr %out0, ptr %out1, ptr %out2, ptr %out3) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load4_v2i32: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld4 { v0.2s, v1.2s, v2.2s, v3.2s }, [x0] ; CHECK-LE-NEXT: str d0, [x1] ; CHECK-LE-NEXT: str d1, [x2] ; CHECK-LE-NEXT: str d2, [x3] ; CHECK-LE-NEXT: str d3, [x4] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load4_v2i32: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld4 { v0.2s, v1.2s, v2.2s, v3.2s }, [x0] ; CHECK-BE-NEXT: st1 { v0.2s }, [x1] ; CHECK-BE-NEXT: st1 { v1.2s }, [x2] ; CHECK-BE-NEXT: st1 { v2.2s }, [x3] ; CHECK-BE-NEXT: st1 { v3.2s }, [x4] ; CHECK-BE-NEXT: ret entry: %loaded = load <8 x i32>, ptr %ptr, align 4, !nontemporal !0 %v0 = shufflevector <8 x i32> %loaded, <8 x i32> poison, <2 x i32> %v1 = shufflevector <8 x i32> %loaded, <8 x i32> poison, <2 x i32> %v2 = shufflevector <8 x i32> %loaded, <8 x i32> poison, <2 x i32> %v3 = shufflevector <8 x i32> %loaded, <8 x i32> poison, <2 x i32> store <2 x i32> %v0, ptr %out0 store <2 x i32> %v1, ptr %out1 store <2 x i32> %v2, ptr %out2 store <2 x i32> %v3, ptr %out3 ret void } define void @test_ldnp_interleaved_load4_v4i16(ptr %ptr, ptr %out0, ptr %out1, ptr %out2, ptr %out3) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load4_v4i16: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld4 { v0.4h, v1.4h, v2.4h, v3.4h }, [x0] ; CHECK-LE-NEXT: str d0, [x1] ; CHECK-LE-NEXT: str d1, [x2] ; CHECK-LE-NEXT: str d2, [x3] ; CHECK-LE-NEXT: str d3, [x4] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load4_v4i16: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld4 { v0.4h, v1.4h, v2.4h, v3.4h }, [x0] ; CHECK-BE-NEXT: st1 { v0.4h }, [x1] ; CHECK-BE-NEXT: st1 { v1.4h }, [x2] ; CHECK-BE-NEXT: st1 { v2.4h }, [x3] ; CHECK-BE-NEXT: st1 { v3.4h }, [x4] ; CHECK-BE-NEXT: ret entry: %loaded = load <16 x i16>, ptr %ptr, align 2, !nontemporal !0 %v0 = shufflevector <16 x i16> %loaded, <16 x i16> poison, <4 x i32> %v1 = shufflevector <16 x i16> %loaded, <16 x i16> poison, <4 x i32> %v2 = shufflevector <16 x i16> %loaded, <16 x i16> poison, <4 x i32> %v3 = shufflevector <16 x i16> %loaded, <16 x i16> poison, <4 x i32> store <4 x i16> %v0, ptr %out0 store <4 x i16> %v1, ptr %out1 store <4 x i16> %v2, ptr %out2 store <4 x i16> %v3, ptr %out3 ret void } define void @test_ldnp_interleaved_load4_v8i8(ptr %ptr, ptr %out0, ptr %out1, ptr %out2, ptr %out3) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load4_v8i8: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld4 { v0.8b, v1.8b, v2.8b, v3.8b }, [x0] ; CHECK-LE-NEXT: str d0, [x1] ; CHECK-LE-NEXT: str d1, [x2] ; CHECK-LE-NEXT: str d2, [x3] ; CHECK-LE-NEXT: str d3, [x4] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load4_v8i8: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld4 { v0.8b, v1.8b, v2.8b, v3.8b }, [x0] ; CHECK-BE-NEXT: st1 { v0.8b }, [x1] ; CHECK-BE-NEXT: st1 { v1.8b }, [x2] ; CHECK-BE-NEXT: st1 { v2.8b }, [x3] ; CHECK-BE-NEXT: st1 { v3.8b }, [x4] ; CHECK-BE-NEXT: ret entry: %loaded = load <32 x i8>, ptr %ptr, align 1, !nontemporal !0 %v0 = shufflevector <32 x i8> %loaded, <32 x i8> poison, <8 x i32> %v1 = shufflevector <32 x i8> %loaded, <32 x i8> poison, <8 x i32> %v2 = shufflevector <32 x i8> %loaded, <32 x i8> poison, <8 x i32> %v3 = shufflevector <32 x i8> %loaded, <32 x i8> poison, <8 x i32> store <8 x i8> %v0, ptr %out0 store <8 x i8> %v1, ptr %out1 store <8 x i8> %v2, ptr %out2 store <8 x i8> %v3, ptr %out3 ret void } define void @test_ldnp_interleaved_load4_v2f32(ptr %ptr, ptr %out0, ptr %out1, ptr %out2, ptr %out3) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load4_v2f32: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld4 { v0.2s, v1.2s, v2.2s, v3.2s }, [x0] ; CHECK-LE-NEXT: str d0, [x1] ; CHECK-LE-NEXT: str d1, [x2] ; CHECK-LE-NEXT: str d2, [x3] ; CHECK-LE-NEXT: str d3, [x4] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load4_v2f32: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld4 { v0.2s, v1.2s, v2.2s, v3.2s }, [x0] ; CHECK-BE-NEXT: st1 { v0.2s }, [x1] ; CHECK-BE-NEXT: st1 { v1.2s }, [x2] ; CHECK-BE-NEXT: st1 { v2.2s }, [x3] ; CHECK-BE-NEXT: st1 { v3.2s }, [x4] ; CHECK-BE-NEXT: ret entry: %loaded = load <8 x float>, ptr %ptr, align 4, !nontemporal !0 %v0 = shufflevector <8 x float> %loaded, <8 x float> poison, <2 x i32> %v1 = shufflevector <8 x float> %loaded, <8 x float> poison, <2 x i32> %v2 = shufflevector <8 x float> %loaded, <8 x float> poison, <2 x i32> %v3 = shufflevector <8 x float> %loaded, <8 x float> poison, <2 x i32> store <2 x float> %v0, ptr %out0 store <2 x float> %v1, ptr %out1 store <2 x float> %v2, ptr %out2 store <2 x float> %v3, ptr %out3 ret void } define void @test_ldnp_interleaved_load4_v4f16(ptr %ptr, ptr %out0, ptr %out1, ptr %out2, ptr %out3) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load4_v4f16: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld4 { v0.4h, v1.4h, v2.4h, v3.4h }, [x0] ; CHECK-LE-NEXT: str d0, [x1] ; CHECK-LE-NEXT: str d1, [x2] ; CHECK-LE-NEXT: str d2, [x3] ; CHECK-LE-NEXT: str d3, [x4] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load4_v4f16: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld4 { v0.4h, v1.4h, v2.4h, v3.4h }, [x0] ; CHECK-BE-NEXT: st1 { v0.4h }, [x1] ; CHECK-BE-NEXT: st1 { v1.4h }, [x2] ; CHECK-BE-NEXT: st1 { v2.4h }, [x3] ; CHECK-BE-NEXT: st1 { v3.4h }, [x4] ; CHECK-BE-NEXT: ret entry: %loaded = load <16 x half>, ptr %ptr, align 2, !nontemporal !0 %v0 = shufflevector <16 x half> %loaded, <16 x half> poison, <4 x i32> %v1 = shufflevector <16 x half> %loaded, <16 x half> poison, <4 x i32> %v2 = shufflevector <16 x half> %loaded, <16 x half> poison, <4 x i32> %v3 = shufflevector <16 x half> %loaded, <16 x half> poison, <4 x i32> store <4 x half> %v0, ptr %out0 store <4 x half> %v1, ptr %out1 store <4 x half> %v2, ptr %out2 store <4 x half> %v3, ptr %out3 ret void } define void @test_ldnp_interleaved_load4_v2i64(ptr %ptr, ptr %out0, ptr %out1, ptr %out2, ptr %out3) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load4_v2i64: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld4 { v0.2d, v1.2d, v2.2d, v3.2d }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: str q2, [x3] ; CHECK-LE-NEXT: str q3, [x4] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load4_v2i64: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld4 { v0.2d, v1.2d, v2.2d, v3.2d }, [x0] ; CHECK-BE-NEXT: st1 { v0.2d }, [x1] ; CHECK-BE-NEXT: st1 { v1.2d }, [x2] ; CHECK-BE-NEXT: st1 { v2.2d }, [x3] ; CHECK-BE-NEXT: st1 { v3.2d }, [x4] ; CHECK-BE-NEXT: ret entry: %loaded = load <8 x i64>, ptr %ptr, align 8, !nontemporal !0 %v0 = shufflevector <8 x i64> %loaded, <8 x i64> poison, <2 x i32> %v1 = shufflevector <8 x i64> %loaded, <8 x i64> poison, <2 x i32> %v2 = shufflevector <8 x i64> %loaded, <8 x i64> poison, <2 x i32> %v3 = shufflevector <8 x i64> %loaded, <8 x i64> poison, <2 x i32> store <2 x i64> %v0, ptr %out0 store <2 x i64> %v1, ptr %out1 store <2 x i64> %v2, ptr %out2 store <2 x i64> %v3, ptr %out3 ret void } define void @test_ldnp_interleaved_load4_v4i32(ptr %ptr, ptr %out0, ptr %out1, ptr %out2, ptr %out3) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load4_v4i32: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld4 { v0.4s, v1.4s, v2.4s, v3.4s }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: str q2, [x3] ; CHECK-LE-NEXT: str q3, [x4] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load4_v4i32: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld4 { v0.4s, v1.4s, v2.4s, v3.4s }, [x0] ; CHECK-BE-NEXT: st1 { v0.4s }, [x1] ; CHECK-BE-NEXT: st1 { v1.4s }, [x2] ; CHECK-BE-NEXT: st1 { v2.4s }, [x3] ; CHECK-BE-NEXT: st1 { v3.4s }, [x4] ; CHECK-BE-NEXT: ret entry: %loaded = load <16 x i32>, ptr %ptr, align 4, !nontemporal !0 %v0 = shufflevector <16 x i32> %loaded, <16 x i32> poison, <4 x i32> %v1 = shufflevector <16 x i32> %loaded, <16 x i32> poison, <4 x i32> %v2 = shufflevector <16 x i32> %loaded, <16 x i32> poison, <4 x i32> %v3 = shufflevector <16 x i32> %loaded, <16 x i32> poison, <4 x i32> store <4 x i32> %v0, ptr %out0 store <4 x i32> %v1, ptr %out1 store <4 x i32> %v2, ptr %out2 store <4 x i32> %v3, ptr %out3 ret void } define void @test_ldnp_interleaved_load4_v8i16(ptr %ptr, ptr %out0, ptr %out1, ptr %out2, ptr %out3) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load4_v8i16: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld4 { v0.8h, v1.8h, v2.8h, v3.8h }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: str q2, [x3] ; CHECK-LE-NEXT: str q3, [x4] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load4_v8i16: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld4 { v0.8h, v1.8h, v2.8h, v3.8h }, [x0] ; CHECK-BE-NEXT: st1 { v0.8h }, [x1] ; CHECK-BE-NEXT: st1 { v1.8h }, [x2] ; CHECK-BE-NEXT: st1 { v2.8h }, [x3] ; CHECK-BE-NEXT: st1 { v3.8h }, [x4] ; CHECK-BE-NEXT: ret entry: %loaded = load <32 x i16>, ptr %ptr, align 2, !nontemporal !0 %v0 = shufflevector <32 x i16> %loaded, <32 x i16> poison, <8 x i32> %v1 = shufflevector <32 x i16> %loaded, <32 x i16> poison, <8 x i32> %v2 = shufflevector <32 x i16> %loaded, <32 x i16> poison, <8 x i32> %v3 = shufflevector <32 x i16> %loaded, <32 x i16> poison, <8 x i32> store <8 x i16> %v0, ptr %out0 store <8 x i16> %v1, ptr %out1 store <8 x i16> %v2, ptr %out2 store <8 x i16> %v3, ptr %out3 ret void } define void @test_ldnp_interleaved_load4_v16i8(ptr %ptr, ptr %out0, ptr %out1, ptr %out2, ptr %out3) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load4_v16i8: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld4 { v0.16b, v1.16b, v2.16b, v3.16b }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: str q2, [x3] ; CHECK-LE-NEXT: str q3, [x4] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load4_v16i8: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld4 { v0.16b, v1.16b, v2.16b, v3.16b }, [x0] ; CHECK-BE-NEXT: st1 { v0.16b }, [x1] ; CHECK-BE-NEXT: st1 { v1.16b }, [x2] ; CHECK-BE-NEXT: st1 { v2.16b }, [x3] ; CHECK-BE-NEXT: st1 { v3.16b }, [x4] ; CHECK-BE-NEXT: ret entry: %loaded = load <64 x i8>, ptr %ptr, align 1, !nontemporal !0 %v0 = shufflevector <64 x i8> %loaded, <64 x i8> poison, <16 x i32> %v1 = shufflevector <64 x i8> %loaded, <64 x i8> poison, <16 x i32> %v2 = shufflevector <64 x i8> %loaded, <64 x i8> poison, <16 x i32> %v3 = shufflevector <64 x i8> %loaded, <64 x i8> poison, <16 x i32> store <16 x i8> %v0, ptr %out0 store <16 x i8> %v1, ptr %out1 store <16 x i8> %v2, ptr %out2 store <16 x i8> %v3, ptr %out3 ret void } define void @test_ldnp_interleaved_load4_v2f64(ptr %ptr, ptr %out0, ptr %out1, ptr %out2, ptr %out3) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load4_v2f64: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld4 { v0.2d, v1.2d, v2.2d, v3.2d }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: str q2, [x3] ; CHECK-LE-NEXT: str q3, [x4] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load4_v2f64: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld4 { v0.2d, v1.2d, v2.2d, v3.2d }, [x0] ; CHECK-BE-NEXT: st1 { v0.2d }, [x1] ; CHECK-BE-NEXT: st1 { v1.2d }, [x2] ; CHECK-BE-NEXT: st1 { v2.2d }, [x3] ; CHECK-BE-NEXT: st1 { v3.2d }, [x4] ; CHECK-BE-NEXT: ret entry: %loaded = load <8 x double>, ptr %ptr, align 8, !nontemporal !0 %v0 = shufflevector <8 x double> %loaded, <8 x double> poison, <2 x i32> %v1 = shufflevector <8 x double> %loaded, <8 x double> poison, <2 x i32> %v2 = shufflevector <8 x double> %loaded, <8 x double> poison, <2 x i32> %v3 = shufflevector <8 x double> %loaded, <8 x double> poison, <2 x i32> store <2 x double> %v0, ptr %out0 store <2 x double> %v1, ptr %out1 store <2 x double> %v2, ptr %out2 store <2 x double> %v3, ptr %out3 ret void } define void @test_ldnp_interleaved_load4_v4f32(ptr %ptr, ptr %out0, ptr %out1, ptr %out2, ptr %out3) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load4_v4f32: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld4 { v0.4s, v1.4s, v2.4s, v3.4s }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: str q2, [x3] ; CHECK-LE-NEXT: str q3, [x4] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load4_v4f32: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld4 { v0.4s, v1.4s, v2.4s, v3.4s }, [x0] ; CHECK-BE-NEXT: st1 { v0.4s }, [x1] ; CHECK-BE-NEXT: st1 { v1.4s }, [x2] ; CHECK-BE-NEXT: st1 { v2.4s }, [x3] ; CHECK-BE-NEXT: st1 { v3.4s }, [x4] ; CHECK-BE-NEXT: ret entry: %loaded = load <16 x float>, ptr %ptr, align 4, !nontemporal !0 %v0 = shufflevector <16 x float> %loaded, <16 x float> poison, <4 x i32> %v1 = shufflevector <16 x float> %loaded, <16 x float> poison, <4 x i32> %v2 = shufflevector <16 x float> %loaded, <16 x float> poison, <4 x i32> %v3 = shufflevector <16 x float> %loaded, <16 x float> poison, <4 x i32> store <4 x float> %v0, ptr %out0 store <4 x float> %v1, ptr %out1 store <4 x float> %v2, ptr %out2 store <4 x float> %v3, ptr %out3 ret void } define void @test_ldnp_interleaved_load4_v8f16(ptr %ptr, ptr %out0, ptr %out1, ptr %out2, ptr %out3) { ; CHECK-LE-LABEL: test_ldnp_interleaved_load4_v8f16: ; CHECK-LE: // %bb.0: // %entry ; CHECK-LE-NEXT: ld4 { v0.8h, v1.8h, v2.8h, v3.8h }, [x0] ; CHECK-LE-NEXT: str q0, [x1] ; CHECK-LE-NEXT: str q1, [x2] ; CHECK-LE-NEXT: str q2, [x3] ; CHECK-LE-NEXT: str q3, [x4] ; CHECK-LE-NEXT: ret ; ; CHECK-BE-LABEL: test_ldnp_interleaved_load4_v8f16: ; CHECK-BE: // %bb.0: // %entry ; CHECK-BE-NEXT: ld4 { v0.8h, v1.8h, v2.8h, v3.8h }, [x0] ; CHECK-BE-NEXT: st1 { v0.8h }, [x1] ; CHECK-BE-NEXT: st1 { v1.8h }, [x2] ; CHECK-BE-NEXT: st1 { v2.8h }, [x3] ; CHECK-BE-NEXT: st1 { v3.8h }, [x4] ; CHECK-BE-NEXT: ret entry: %loaded = load <32 x half>, ptr %ptr, align 2, !nontemporal !0 %v0 = shufflevector <32 x half> %loaded, <32 x half> poison, <8 x i32> %v1 = shufflevector <32 x half> %loaded, <32 x half> poison, <8 x i32> %v2 = shufflevector <32 x half> %loaded, <32 x half> poison, <8 x i32> %v3 = shufflevector <32 x half> %loaded, <32 x half> poison, <8 x i32> store <8 x half> %v0, ptr %out0 store <8 x half> %v1, ptr %out1 store <8 x half> %v2, ptr %out2 store <8 x half> %v3, ptr %out3 ret void } !0 = !{ i32 1 }