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:   -mtriple=powerpc64le-unknown-linux-gnu -mcpu=pwr9 < %s | FileCheck %s
4
5; FIXME: PPCLoopInstrFormPrep should be able to common base for "(unsigned long long *)(p + j + 5)"
6; and "(unsigned long long *)(p + j + 9)", thus we only have two DS form load inside the loop.
7
8; long long foo(char *p, int n, int count) {
9;   int j = 0;
10;   long long sum = 0;
11;   for (int i = 0; i < n; i++) {
12;     sum += *(unsigned long long *)(p + j + 5);
13;     sum += *(unsigned long long *)(p + j + 9);
14;     j += count;
15;   }
16;   return sum;
17; }
18
19define i64 @foo(i8* %p, i32 signext %n, i32 signext %count) {
20; CHECK-LABEL: foo:
21; CHECK:       # %bb.0: # %entry
22; CHECK-NEXT:    cmpwi r4, 1
23; CHECK-NEXT:    blt cr0, .LBB0_4
24; CHECK-NEXT:  # %bb.1: # %for.body.preheader
25; CHECK-NEXT:    addi r4, r4, -1
26; CHECK-NEXT:    extsw r5, r5
27; CHECK-NEXT:    li r6, 0
28; CHECK-NEXT:    li r7, 5
29; CHECK-NEXT:    li r8, 9
30; CHECK-NEXT:    clrldi r4, r4, 32
31; CHECK-NEXT:    addi r4, r4, 1
32; CHECK-NEXT:    mtctr r4
33; CHECK-NEXT:    li r4, 0
34; CHECK-NEXT:    .p2align 5
35; CHECK-NEXT:  .LBB0_2: # %for.body
36; CHECK-NEXT:    #
37; CHECK-NEXT:    add r9, r3, r6
38; CHECK-NEXT:    add r6, r6, r5
39; CHECK-NEXT:    ldx r10, r9, r7
40; CHECK-NEXT:    ldx r9, r9, r8
41; CHECK-NEXT:    add r4, r10, r4
42; CHECK-NEXT:    add r4, r4, r9
43; CHECK-NEXT:    bdnz .LBB0_2
44; CHECK-NEXT:  # %bb.3: # %for.cond.cleanup
45; CHECK-NEXT:    mr r3, r4
46; CHECK-NEXT:    blr
47; CHECK-NEXT:  .LBB0_4:
48; CHECK-NEXT:    li r4, 0
49; CHECK-NEXT:    mr r3, r4
50; CHECK-NEXT:    blr
51entry:
52  %cmp16 = icmp sgt i32 %n, 0
53  br i1 %cmp16, label %for.body.preheader, label %for.cond.cleanup
54
55for.body.preheader:                               ; preds = %entry
56  %0 = sext i32 %count to i64
57  br label %for.body
58
59for.cond.cleanup:                                 ; preds = %for.body, %entry
60  %sum.0.lcssa = phi i64 [ 0, %entry ], [ %add5, %for.body ]
61  ret i64 %sum.0.lcssa
62
63for.body:                                         ; preds = %for.body.preheader, %for.body
64  %indvars.iv = phi i64 [ 0, %for.body.preheader ], [ %indvars.iv.next, %for.body ]
65  %i.019 = phi i32 [ 0, %for.body.preheader ], [ %inc, %for.body ]
66  %sum.018 = phi i64 [ 0, %for.body.preheader ], [ %add5, %for.body ]
67  %add.ptr = getelementptr inbounds i8, i8* %p, i64 %indvars.iv
68  %add.ptr1 = getelementptr inbounds i8, i8* %add.ptr, i64 5
69  %1 = bitcast i8* %add.ptr1 to i64*
70  %2 = load i64, i64* %1, align 8
71  %add = add i64 %2, %sum.018
72  %add.ptr4 = getelementptr inbounds i8, i8* %add.ptr, i64 9
73  %3 = bitcast i8* %add.ptr4 to i64*
74  %4 = load i64, i64* %3, align 8
75  %add5 = add i64 %add, %4
76  %indvars.iv.next = add nsw i64 %indvars.iv, %0
77  %inc = add nuw nsw i32 %i.019, 1
78  %exitcond.not = icmp eq i32 %inc, %n
79  br i1 %exitcond.not, label %for.cond.cleanup, label %for.body
80}
81
82; char foo1(char *p, int n, int count) {
83;   int j = 0;
84;   char sum = 0;
85;   for (int i = 0; i < n; i++) {
86;     sum += *(p + j + 1000);
87;     j += count;
88;    }
89;   return sum;
90; }
91
92define zeroext i8 @foo1(i8* %p, i32 signext %n, i32 signext %count) {
93; CHECK-LABEL: foo1:
94; CHECK:       # %bb.0: # %entry
95; CHECK-NEXT:    cmpwi r4, 1
96; CHECK-NEXT:    blt cr0, .LBB1_4
97; CHECK-NEXT:  # %bb.1: # %for.body.preheader
98; CHECK-NEXT:    addi r4, r4, -1
99; CHECK-NEXT:    addi r3, r3, 1000
100; CHECK-NEXT:    extsw r5, r5
101; CHECK-NEXT:    li r6, 0
102; CHECK-NEXT:    clrldi r4, r4, 32
103; CHECK-NEXT:    addi r4, r4, 1
104; CHECK-NEXT:    mtctr r4
105; CHECK-NEXT:    li r4, 0
106; CHECK-NEXT:    .p2align 4
107; CHECK-NEXT:  .LBB1_2: # %for.body
108; CHECK-NEXT:    #
109; CHECK-NEXT:    lbzx r7, r3, r6
110; CHECK-NEXT:    add r6, r6, r5
111; CHECK-NEXT:    add r4, r7, r4
112; CHECK-NEXT:    bdnz .LBB1_2
113; CHECK-NEXT:  # %bb.3: # %for.cond.cleanup
114; CHECK-NEXT:    clrldi r3, r4, 56
115; CHECK-NEXT:    blr
116; CHECK-NEXT:  .LBB1_4:
117; CHECK-NEXT:    li r3, 0
118; CHECK-NEXT:    blr
119entry:
120  %cmp10 = icmp sgt i32 %n, 0
121  br i1 %cmp10, label %for.body.preheader, label %for.cond.cleanup
122
123for.body.preheader:                               ; preds = %entry
124  %0 = sext i32 %count to i64
125  %add.ptr = getelementptr inbounds i8, i8* %p, i64 1000
126  br label %for.body
127
128for.cond.cleanup:                                 ; preds = %for.body, %entry
129  %sum.0.lcssa = phi i8 [ 0, %entry ], [ %add, %for.body ]
130  ret i8 %sum.0.lcssa
131
132for.body:                                         ; preds = %for.body.preheader, %for.body
133  %indvars.iv = phi i64 [ 0, %for.body.preheader ], [ %indvars.iv.next, %for.body ]
134  %i.013 = phi i32 [ 0, %for.body.preheader ], [ %inc, %for.body ]
135  %sum.012 = phi i8 [ 0, %for.body.preheader ], [ %add, %for.body ]
136  %add.ptr1 = getelementptr inbounds i8, i8* %add.ptr, i64 %indvars.iv
137  %1 = load i8, i8* %add.ptr1, align 1
138  %add = add i8 %1, %sum.012
139  %indvars.iv.next = add nsw i64 %indvars.iv, %0
140  %inc = add nuw nsw i32 %i.013, 1
141  %exitcond.not = icmp eq i32 %inc, %n
142  br i1 %exitcond.not, label %for.cond.cleanup, label %for.body
143}
144