1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -O0 < %s | FileCheck %s
3 target datalayout = "e-m:e-i64:64-n32:64"
4 target triple = "powerpc64le-unknown-linux-gnu"
6 define void @bn_mul_comba8(i64* nocapture %r, i64* nocapture readonly %a, i64* nocapture readonly %b) {
7 ; CHECK-LABEL: bn_mul_comba8:
9 ; CHECK-NEXT: ld 6, 0(4)
10 ; CHECK-NEXT: ld 7, 0(5)
11 ; CHECK-NEXT: mulhdu 8, 7, 6
12 ; CHECK-NEXT: ld 4, 8(4)
13 ; CHECK-NEXT: mulld 9, 4, 6
14 ; CHECK-NEXT: mulhdu 4, 4, 6
15 ; CHECK-NEXT: addc 6, 9, 8
16 ; CHECK-NEXT: addze 4, 4
17 ; CHECK-NEXT: ld 5, 8(5)
18 ; CHECK-NEXT: mulld 8, 5, 7
19 ; CHECK-NEXT: mulhdu 5, 5, 7
20 ; CHECK-NEXT: addc 6, 6, 8
21 ; CHECK-NEXT: addze 5, 5
22 ; CHECK-NEXT: add 4, 5, 4
23 ; CHECK-NEXT: cmpld 7, 4, 5
24 ; CHECK-NEXT: mfocrf 10, 1
25 ; CHECK-NEXT: rlwinm 10, 10, 29, 31, 31
26 ; CHECK-NEXT: # implicit-def: $x4
27 ; CHECK-NEXT: mr 4, 10
28 ; CHECK-NEXT: clrldi 4, 4, 32
29 ; CHECK-NEXT: std 4, 0(3)
30 ; CHECK-NEXT: std 6, -8(1) # 8-byte Folded Spill
32 %1 = load i64, i64* %a, align 8
33 %conv = zext i64 %1 to i128
34 %2 = load i64, i64* %b, align 8
35 %conv2 = zext i64 %2 to i128
36 %mul = mul nuw i128 %conv2, %conv
37 %shr = lshr i128 %mul, 64
38 %agep = getelementptr inbounds i64, i64* %a, i64 1
39 %3 = load i64, i64* %agep, align 8
40 %conv14 = zext i64 %3 to i128
41 %mul15 = mul nuw i128 %conv14, %conv
42 %add17 = add i128 %mul15, %shr
43 %shr19 = lshr i128 %add17, 64
44 %conv20 = trunc i128 %shr19 to i64
45 %bgep = getelementptr inbounds i64, i64* %b, i64 1
46 %4 = load i64, i64* %bgep, align 8
47 %conv28 = zext i64 %4 to i128
48 %mul31 = mul nuw i128 %conv28, %conv2
49 %conv32 = and i128 %add17, 18446744073709551615
50 %add33 = add i128 %conv32, %mul31
51 %shr35 = lshr i128 %add33, 64
52 %conv36 = trunc i128 %shr35 to i64
53 %add37 = add i64 %conv36, %conv20
54 %cmp38 = icmp ult i64 %add37, %conv36
55 %conv148 = zext i1 %cmp38 to i64
56 store i64 %conv148, i64* %r, align 8