89 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; Test high-part vector multiplication.3;4; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z13 | FileCheck %s5 6; Test a v16i8 unsigned high-part multiplication.7define <16 x i8> @f1(<16 x i8> %val1, <16 x i8> %val2) {8; CHECK-LABEL: f1:9; CHECK: # %bb.0:10; CHECK-NEXT: vmlhb %v24, %v24, %v2611; CHECK-NEXT: br %r1412 %zext1 = zext <16 x i8> %val1 to <16 x i16>13 %zext2 = zext <16 x i8> %val2 to <16 x i16>14 %mulx = mul <16 x i16> %zext1, %zext215 %highx = lshr <16 x i16> %mulx, splat(i16 8)16 %high = trunc <16 x i16> %highx to <16 x i8>17 ret <16 x i8> %high18}19 20; Test a v16i8 signed high-part multiplication.21define <16 x i8> @f2(<16 x i8> %val1, <16 x i8> %val2) {22; CHECK-LABEL: f2:23; CHECK: # %bb.0:24; CHECK-NEXT: vmhb %v24, %v24, %v2625; CHECK-NEXT: br %r1426 %sext1 = sext <16 x i8> %val1 to <16 x i16>27 %sext2 = sext <16 x i8> %val2 to <16 x i16>28 %mulx = mul <16 x i16> %sext1, %sext229 %highx = lshr <16 x i16> %mulx, splat(i16 8)30 %high = trunc <16 x i16> %highx to <16 x i8>31 ret <16 x i8> %high32}33 34; Test a v8i16 unsigned high-part multiplication.35define <8 x i16> @f3(<8 x i16> %val1, <8 x i16> %val2) {36; CHECK-LABEL: f3:37; CHECK: # %bb.0:38; CHECK-NEXT: vmlhh %v24, %v24, %v2639; CHECK-NEXT: br %r1440 %zext1 = zext <8 x i16> %val1 to <8 x i32>41 %zext2 = zext <8 x i16> %val2 to <8 x i32>42 %mulx = mul <8 x i32> %zext1, %zext243 %highx = lshr <8 x i32> %mulx, splat(i32 16)44 %high = trunc <8 x i32> %highx to <8 x i16>45 ret <8 x i16> %high46}47 48; Test a v8i16 signed high-part multiplication.49define <8 x i16> @f4(<8 x i16> %val1, <8 x i16> %val2) {50; CHECK-LABEL: f4:51; CHECK: # %bb.0:52; CHECK-NEXT: vmhh %v24, %v24, %v2653; CHECK-NEXT: br %r1454 %sext1 = sext <8 x i16> %val1 to <8 x i32>55 %sext2 = sext <8 x i16> %val2 to <8 x i32>56 %mulx = mul <8 x i32> %sext1, %sext257 %highx = lshr <8 x i32> %mulx, splat(i32 16)58 %high = trunc <8 x i32> %highx to <8 x i16>59 ret <8 x i16> %high60}61 62; Test a v4i32 unsigned high-part multiplication.63define <4 x i32> @f5(<4 x i32> %val1, <4 x i32> %val2) {64; CHECK-LABEL: f5:65; CHECK: # %bb.0:66; CHECK-NEXT: vmlhf %v24, %v24, %v2667; CHECK-NEXT: br %r1468 %zext1 = zext <4 x i32> %val1 to <4 x i64>69 %zext2 = zext <4 x i32> %val2 to <4 x i64>70 %mulx = mul <4 x i64> %zext1, %zext271 %highx = lshr <4 x i64> %mulx, splat(i64 32)72 %high = trunc <4 x i64> %highx to <4 x i32>73 ret <4 x i32> %high74}75 76; Test a v4i32 signed high-part multiplication.77define <4 x i32> @f6(<4 x i32> %val1, <4 x i32> %val2) {78; CHECK-LABEL: f6:79; CHECK: # %bb.0:80; CHECK-NEXT: vmhf %v24, %v24, %v2681; CHECK-NEXT: br %r1482 %sext1 = sext <4 x i32> %val1 to <4 x i64>83 %sext2 = sext <4 x i32> %val2 to <4 x i64>84 %mulx = mul <4 x i64> %sext1, %sext285 %highx = lshr <4 x i64> %mulx, splat(i64 32)86 %high = trunc <4 x i64> %highx to <4 x i32>87 ret <4 x i32> %high88}89