85 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s --mtriple=x86_64-unknown-linux-gnu -mattr=+sse4.1 | FileCheck %s3 4define void @pmuldq(ptr nocapture %0, i32 %1, i64 %2) {5; CHECK-LABEL: pmuldq:6; CHECK: # %bb.0:7; CHECK-NEXT: testq %rdx, %rdx8; CHECK-NEXT: je .LBB0_39; CHECK-NEXT: # %bb.1:10; CHECK-NEXT: movd %esi, %xmm011; CHECK-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]12; CHECK-NEXT: .p2align 413; CHECK-NEXT: .LBB0_2: # =>This Inner Loop Header: Depth=114; CHECK-NEXT: movdqa (%rdi), %xmm115; CHECK-NEXT: pmuldq %xmm0, %xmm116; CHECK-NEXT: movdqa %xmm1, (%rdi)17; CHECK-NEXT: addq $16, %rdi18; CHECK-NEXT: decq %rdx19; CHECK-NEXT: jne .LBB0_220; CHECK-NEXT: .LBB0_3:21; CHECK-NEXT: retq22 %4 = insertelement <4 x i32> undef, i32 %1, i32 023 %5 = shufflevector <4 x i32> %4, <4 x i32> poison, <4 x i32> zeroinitializer24 %6 = bitcast <4 x i32> %5 to <2 x i64>25 %7 = shl <2 x i64> %6, <i64 32, i64 32>26 %8 = ashr exact <2 x i64> %7, <i64 32, i64 32>27 %9 = icmp eq i64 %2, 028 br i1 %9, label %10, label %1129 3010: ; preds = %11, %331 ret void32 3311: ; preds = %3, %1134 %12 = phi i64 [ %18, %11 ], [ 0, %3 ]35 %13 = getelementptr inbounds <2 x i64>, ptr %0, i64 %1236 %14 = load <2 x i64>, ptr %13, align 1637 %15 = shl <2 x i64> %14, <i64 32, i64 32>38 %16 = ashr exact <2 x i64> %15, <i64 32, i64 32>39 %17 = mul nsw <2 x i64> %16, %840 store <2 x i64> %17, ptr %13, align 1641 %18 = add nuw i64 %12, 142 %19 = icmp eq i64 %18, %243 br i1 %19, label %10, label %1144}45 46define void @pmuludq(ptr nocapture %0, i32 %1, i64 %2) {47; CHECK-LABEL: pmuludq:48; CHECK: # %bb.0:49; CHECK-NEXT: testq %rdx, %rdx50; CHECK-NEXT: je .LBB1_351; CHECK-NEXT: # %bb.1:52; CHECK-NEXT: movd %esi, %xmm053; CHECK-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]54; CHECK-NEXT: .p2align 455; CHECK-NEXT: .LBB1_2: # =>This Inner Loop Header: Depth=156; CHECK-NEXT: movdqa (%rdi), %xmm157; CHECK-NEXT: pmuludq %xmm0, %xmm158; CHECK-NEXT: movdqa %xmm1, (%rdi)59; CHECK-NEXT: addq $16, %rdi60; CHECK-NEXT: decq %rdx61; CHECK-NEXT: jne .LBB1_262; CHECK-NEXT: .LBB1_3:63; CHECK-NEXT: retq64 %4 = insertelement <4 x i32> undef, i32 %1, i32 065 %5 = shufflevector <4 x i32> %4, <4 x i32> poison, <4 x i32> zeroinitializer66 %6 = bitcast <4 x i32> %5 to <2 x i64>67 %7 = and <2 x i64> %6, <i64 4294967295, i64 4294967295>68 %8 = icmp eq i64 %2, 069 br i1 %8, label %9, label %1070 719: ; preds = %10, %372 ret void73 7410: ; preds = %3, %1075 %11 = phi i64 [ %16, %10 ], [ 0, %3 ]76 %12 = getelementptr inbounds <2 x i64>, ptr %0, i64 %1177 %13 = load <2 x i64>, ptr %12, align 1678 %14 = and <2 x i64> %13, <i64 4294967295, i64 4294967295>79 %15 = mul nuw <2 x i64> %14, %780 store <2 x i64> %15, ptr %12, align 1681 %16 = add nuw i64 %11, 182 %17 = icmp eq i64 %16, %283 br i1 %17, label %9, label %1084}85