75 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 42; RUN: llc -mtriple=aarch64-linux -mattr=+neon,+fp8dot2,+fp8dot4 < %s | FileCheck %s3 4define <4 x half> @test_fdot_f16(<4 x half> %vd, <8 x i8> %vn, <8 x i8> %vm) {5; CHECK-LABEL: test_fdot_f16:6; CHECK: // %bb.0:7; CHECK-NEXT: fdot v0.4h, v1.8b, v2.8b8; CHECK-NEXT: ret9 %res = call <4 x half> @llvm.aarch64.neon.fp8.fdot2.v4f16.v8i8(<4 x half> %vd, <8 x i8> %vn, <8 x i8> %vm)10 ret <4 x half> %res11}12 13define <8 x half> @test_fdotq_f16(<8 x half> %vd, <16 x i8> %vn, <16 x i8> %vm) {14; CHECK-LABEL: test_fdotq_f16:15; CHECK: // %bb.0:16; CHECK-NEXT: fdot v0.8h, v1.16b, v2.16b17; CHECK-NEXT: ret18 %res = call <8 x half> @llvm.aarch64.neon.fp8.fdot2.v8f16.v16i8(<8 x half> %vd, <16 x i8> %vn, <16 x i8> %vm)19 ret <8 x half> %res20}21 22define <4 x half> @test_fdot_lane_f16(<4 x half> %vd, <8 x i8> %vn, <16 x i8> %vm) {23; CHECK-LABEL: test_fdot_lane_f16:24; CHECK: // %bb.0:25; CHECK-NEXT: fdot v0.4h, v1.8b, v2.2b[0]26; CHECK-NEXT: ret27 %res = call <4 x half> @llvm.aarch64.neon.fp8.fdot2.lane.v4f16.v8i8(<4 x half> %vd, <8 x i8> %vn, <16 x i8> %vm, i32 0)28 ret <4 x half> %res29}30 31define <8 x half> @test_fdotq_lane_f16(<8 x half> %vd, <16 x i8> %vn, <16 x i8> %vm) {32; CHECK-LABEL: test_fdotq_lane_f16:33; CHECK: // %bb.0:34; CHECK-NEXT: fdot v0.8h, v1.16b, v2.2b[7]35; CHECK-NEXT: ret36 %res = call <8 x half> @llvm.aarch64.neon.fp8.fdot2.lane.v8f16.v16i8(<8 x half> %vd, <16 x i8> %vn, <16 x i8> %vm, i32 7)37 ret <8 x half> %res38}39 40define <2 x float> @test_fdot_f32(<2 x float> %vd, <8 x i8> %vn, <8 x i8> %vm) {41; CHECK-LABEL: test_fdot_f32:42; CHECK: // %bb.0:43; CHECK-NEXT: fdot v0.2s, v1.8b, v2.8b44; CHECK-NEXT: ret45 %res = call <2 x float> @llvm.aarch64.neon.fp8.fdot4.v2f32.v8i8(<2 x float> %vd, <8 x i8> %vn, <8 x i8> %vm)46 ret <2 x float> %res47}48 49define <4 x float> @test_fdotq_f32(<4 x float> %vd, <16 x i8> %vn, <16 x i8> %vm) {50; CHECK-LABEL: test_fdotq_f32:51; CHECK: // %bb.0:52; CHECK-NEXT: fdot v0.4s, v1.16b, v2.16b53; CHECK-NEXT: ret54 %res = call <4 x float> @llvm.aarch64.neon.fp8.fdot4.v4f32.v16i8(<4 x float> %vd, <16 x i8> %vn, <16 x i8> %vm)55 ret <4 x float> %res56}57 58define <2 x float> @test_fdot_lane_f32(<2 x float> %vd, <8 x i8> %vn, <16 x i8> %vm) {59; CHECK-LABEL: test_fdot_lane_f32:60; CHECK: // %bb.0:61; CHECK-NEXT: fdot v0.2s, v1.8b, v2.4b[0]62; CHECK-NEXT: ret63 %res = call <2 x float> @llvm.aarch64.neon.fp8.fdot4.lane.v2f32.v8i8(<2 x float> %vd, <8 x i8> %vn, <16 x i8> %vm, i32 0)64 ret <2 x float> %res65}66 67define <4 x float> @test_fdotq_lane_f32(<4 x float> %vd, <16 x i8> %vn, <16 x i8> %vm) {68; CHECK-LABEL: test_fdotq_lane_f32:69; CHECK: // %bb.0:70; CHECK-NEXT: fdot v0.4s, v1.16b, v2.4b[3]71; CHECK-NEXT: ret72 %res = call <4 x float> @llvm.aarch64.neon.fp8.fdot4.lane.v4f32.v16i8(<4 x float> %vd, <16 x i8> %vn, <16 x i8> %vm, i32 3)73 ret <4 x float> %res74}75