1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -disable-lsr -ppc-asm-full-reg-names -verify-machineinstrs \
3; RUN:   -ppc-formprep-update-nonconst-inc -mtriple=powerpc64le-unknown-linux-gnu \
4; RUN:   -mcpu=pwr9 < %s | FileCheck %s
5
6; long long foo(char *p, int n, int count) {
7;   int j = 0;
8;   long long sum = 0;
9;   for (int i = 0; i < n; i++) {
10;     sum += *(unsigned long long *)(p + j + 5);
11;     sum += *(unsigned long long *)(p + j + 9);
12;     j += count;
13;   }
14;   return sum;
15; }
16
17define i64 @foo(i8* %p, i32 signext %n, i32 signext %count) {
18; CHECK-LABEL: foo:
19; CHECK:       # %bb.0: # %entry
20; CHECK-NEXT:    cmpwi r4, 1
21; CHECK-NEXT:    blt cr0, .LBB0_4
22; CHECK-NEXT:  # %bb.1: # %for.body.preheader
23; CHECK-NEXT:    addi r6, r3, 5
24; CHECK-NEXT:    addi r3, r4, -1
25; CHECK-NEXT:    extsw r5, r5
26; CHECK-NEXT:    clrldi r3, r3, 32
27; CHECK-NEXT:    addi r3, r3, 1
28; CHECK-NEXT:    mtctr r3
29; CHECK-NEXT:    li r3, 0
30; CHECK-NEXT:    .p2align 5
31; CHECK-NEXT:  .LBB0_2: # %for.body
32; CHECK-NEXT:    #
33; CHECK-NEXT:    ld r4, 0(r6)
34; CHECK-NEXT:    add r3, r4, r3
35; CHECK-NEXT:    ld r4, 4(r6)
36; CHECK-NEXT:    add r6, r6, r5
37; CHECK-NEXT:    add r3, r3, r4
38; CHECK-NEXT:    bdnz .LBB0_2
39; CHECK-NEXT:  # %bb.3: # %for.cond.cleanup
40; CHECK-NEXT:    blr
41; CHECK-NEXT:  .LBB0_4:
42; CHECK-NEXT:    li r3, 0
43; CHECK-NEXT:    blr
44entry:
45  %cmp16 = icmp sgt i32 %n, 0
46  br i1 %cmp16, label %for.body.preheader, label %for.cond.cleanup
47
48for.body.preheader:                               ; preds = %entry
49  %0 = sext i32 %count to i64
50  br label %for.body
51
52for.cond.cleanup:                                 ; preds = %for.body, %entry
53  %sum.0.lcssa = phi i64 [ 0, %entry ], [ %add5, %for.body ]
54  ret i64 %sum.0.lcssa
55
56for.body:                                         ; preds = %for.body.preheader, %for.body
57  %indvars.iv = phi i64 [ 0, %for.body.preheader ], [ %indvars.iv.next, %for.body ]
58  %i.019 = phi i32 [ 0, %for.body.preheader ], [ %inc, %for.body ]
59  %sum.018 = phi i64 [ 0, %for.body.preheader ], [ %add5, %for.body ]
60  %add.ptr = getelementptr inbounds i8, i8* %p, i64 %indvars.iv
61  %add.ptr1 = getelementptr inbounds i8, i8* %add.ptr, i64 5
62  %1 = bitcast i8* %add.ptr1 to i64*
63  %2 = load i64, i64* %1, align 8
64  %add = add i64 %2, %sum.018
65  %add.ptr4 = getelementptr inbounds i8, i8* %add.ptr, i64 9
66  %3 = bitcast i8* %add.ptr4 to i64*
67  %4 = load i64, i64* %3, align 8
68  %add5 = add i64 %add, %4
69  %indvars.iv.next = add nsw i64 %indvars.iv, %0
70  %inc = add nuw nsw i32 %i.019, 1
71  %exitcond.not = icmp eq i32 %inc, %n
72  br i1 %exitcond.not, label %for.cond.cleanup, label %for.body
73}
74
75; char foo1(char *p, int n, int count) {
76;   int j = 0;
77;   char sum = 0;
78;   for (int i = 0; i < n; i++) {
79;     sum += *(p + j + 1000);
80;     j += count;
81;    }
82;   return sum;
83; }
84
85define zeroext i8 @foo1(i8* %p, i32 signext %n, i32 signext %count) {
86; CHECK-LABEL: foo1:
87; CHECK:       # %bb.0: # %entry
88; CHECK-NEXT:    cmpwi r4, 1
89; CHECK-NEXT:    blt cr0, .LBB1_4
90; CHECK-NEXT:  # %bb.1: # %for.body.preheader
91; CHECK-NEXT:    extsw r5, r5
92; CHECK-NEXT:    sub r3, r3, r5
93; CHECK-NEXT:    addi r6, r3, 1000
94; CHECK-NEXT:    addi r3, r4, -1
95; CHECK-NEXT:    clrldi r3, r3, 32
96; CHECK-NEXT:    addi r3, r3, 1
97; CHECK-NEXT:    mtctr r3
98; CHECK-NEXT:    li r3, 0
99; CHECK-NEXT:    .p2align 4
100; CHECK-NEXT:  .LBB1_2: # %for.body
101; CHECK-NEXT:    #
102; CHECK-NEXT:    lbzux r4, r6, r5
103; CHECK-NEXT:    add r3, r4, r3
104; CHECK-NEXT:    bdnz .LBB1_2
105; CHECK-NEXT:  # %bb.3: # %for.cond.cleanup
106; CHECK-NEXT:    clrldi r3, r3, 56
107; CHECK-NEXT:    blr
108; CHECK-NEXT:  .LBB1_4:
109; CHECK-NEXT:    li r3, 0
110; CHECK-NEXT:    blr
111entry:
112  %cmp10 = icmp sgt i32 %n, 0
113  br i1 %cmp10, label %for.body.preheader, label %for.cond.cleanup
114
115for.body.preheader:                               ; preds = %entry
116  %0 = sext i32 %count to i64
117  %add.ptr = getelementptr inbounds i8, i8* %p, i64 1000
118  br label %for.body
119
120for.cond.cleanup:                                 ; preds = %for.body, %entry
121  %sum.0.lcssa = phi i8 [ 0, %entry ], [ %add, %for.body ]
122  ret i8 %sum.0.lcssa
123
124for.body:                                         ; preds = %for.body.preheader, %for.body
125  %indvars.iv = phi i64 [ 0, %for.body.preheader ], [ %indvars.iv.next, %for.body ]
126  %i.013 = phi i32 [ 0, %for.body.preheader ], [ %inc, %for.body ]
127  %sum.012 = phi i8 [ 0, %for.body.preheader ], [ %add, %for.body ]
128  %add.ptr1 = getelementptr inbounds i8, i8* %add.ptr, i64 %indvars.iv
129  %1 = load i8, i8* %add.ptr1, align 1
130  %add = add i8 %1, %sum.012
131  %indvars.iv.next = add nsw i64 %indvars.iv, %0
132  %inc = add nuw nsw i32 %i.013, 1
133  %exitcond.not = icmp eq i32 %inc, %n
134  br i1 %exitcond.not, label %for.cond.cleanup, label %for.body
135}
136