95 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -mtriple aarch64-none-linux-gnu -mattr=+dotprod < %s | FileCheck %s3 4define fastcc void @test_sdot_v4i8(ptr noalias nocapture %0, ptr noalias nocapture readonly %1, ptr noalias nocapture readonly %2) {5; CHECK-LABEL: test_sdot_v4i8:6; CHECK: // %bb.0: // %entry7; CHECK-NEXT: ldr w8, [x2]8; CHECK-NEXT: ldr w9, [x1]9; CHECK-NEXT: dup v0.2s, wzr10; CHECK-NEXT: fmov s1, w811; CHECK-NEXT: fmov s2, w912; CHECK-NEXT: sdot v0.2s, v1.8b, v2.8b13; CHECK-NEXT: fmov x8, d014; CHECK-NEXT: str w8, [x0]15; CHECK-NEXT: ret16entry:17 %3 = load i8, ptr %1, align 118 %4 = sext i8 %3 to i3219 %5 = load i8, ptr %2, align 120 %6 = sext i8 %5 to i3221 %7 = mul nsw i32 %6, %422 %8 = getelementptr inbounds i8, ptr %1, i64 123 %9 = load i8, ptr %8, align 124 %10 = sext i8 %9 to i3225 %11 = getelementptr inbounds i8, ptr %2, i64 126 %12 = load i8, ptr %11, align 127 %13 = sext i8 %12 to i3228 %14 = mul nsw i32 %13, %1029 %15 = add nsw i32 %14, %730 %16 = getelementptr inbounds i8, ptr %1, i64 231 %17 = load i8, ptr %16, align 132 %18 = sext i8 %17 to i3233 %19 = getelementptr inbounds i8, ptr %2, i64 234 %20 = load i8, ptr %19, align 135 %21 = sext i8 %20 to i3236 %22 = mul nsw i32 %21, %1837 %23 = add nsw i32 %22, %1538 %24 = getelementptr inbounds i8, ptr %1, i64 339 %25 = load i8, ptr %24, align 140 %26 = sext i8 %25 to i3241 %27 = getelementptr inbounds i8, ptr %2, i64 342 %28 = load i8, ptr %27, align 143 %29 = sext i8 %28 to i3244 %30 = mul nsw i32 %29, %2645 %31 = add nsw i32 %30, %2346 store i32 %31, ptr %0, align 6447 ret void48}49 50define fastcc void @test_udot_v4i8(ptr noalias nocapture %0, ptr noalias nocapture readonly %1, ptr noalias nocapture readonly %2) {51; CHECK-LABEL: test_udot_v4i8:52; CHECK: // %bb.0: // %entry53; CHECK-NEXT: ldr w8, [x2]54; CHECK-NEXT: ldr w9, [x1]55; CHECK-NEXT: dup v0.2s, wzr56; CHECK-NEXT: fmov s1, w857; CHECK-NEXT: fmov s2, w958; CHECK-NEXT: udot v0.2s, v1.8b, v2.8b59; CHECK-NEXT: fmov x8, d060; CHECK-NEXT: str w8, [x0]61; CHECK-NEXT: ret62entry:63 %3 = load i8, ptr %1, align 164 %4 = zext i8 %3 to i3265 %5 = load i8, ptr %2, align 166 %6 = zext i8 %5 to i3267 %7 = mul nsw i32 %6, %468 %8 = getelementptr inbounds i8, ptr %1, i64 169 %9 = load i8, ptr %8, align 170 %10 = zext i8 %9 to i3271 %11 = getelementptr inbounds i8, ptr %2, i64 172 %12 = load i8, ptr %11, align 173 %13 = zext i8 %12 to i3274 %14 = mul nsw i32 %13, %1075 %15 = add nsw i32 %14, %776 %16 = getelementptr inbounds i8, ptr %1, i64 277 %17 = load i8, ptr %16, align 178 %18 = zext i8 %17 to i3279 %19 = getelementptr inbounds i8, ptr %2, i64 280 %20 = load i8, ptr %19, align 181 %21 = zext i8 %20 to i3282 %22 = mul nsw i32 %21, %1883 %23 = add nsw i32 %22, %1584 %24 = getelementptr inbounds i8, ptr %1, i64 385 %25 = load i8, ptr %24, align 186 %26 = zext i8 %25 to i3287 %27 = getelementptr inbounds i8, ptr %2, i64 388 %28 = load i8, ptr %27, align 189 %29 = zext i8 %28 to i3290 %30 = mul nsw i32 %29, %2691 %31 = add nsw i32 %30, %2392 store i32 %31, ptr %0, align 6493 ret void94}95