xref: /llvm-project/llvm/test/Transforms/SLPVectorizer/X86/external_user.ll (revision 38fffa630ee80163dc65e759392ad29798905679)
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt < %s -passes=slp-vectorizer,dce -S -mtriple=x86_64-apple-macosx10.8.0 -mcpu=corei7-avx | FileCheck %s
3
4target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
5target triple = "x86_64-apple-macosx10.8.0"
6
7; double foo(ptr restrict b,  ptr restrict a, int n, int m) {
8;   double r=a[1];
9;   double g=a[0];
10;   double x;
11;   for (int i=0; i < 100; i++) {
12;     r += 10;
13;     g += 10;
14;     r *= 4;
15;     g *= 4;
16;     x = g; <----- external user!
17;     r += 4;
18;     g += 4;
19;   }
20;   b[0] = g;
21;   b[1] = r;
22;
23;   return x; <-- must extract here!
24; }
25
26define double @ext_user(ptr noalias nocapture %B, ptr noalias nocapture %A, i32 %n, i32 %m) {
27; CHECK-LABEL: @ext_user(
28; CHECK-NEXT:  entry:
29; CHECK-NEXT:    [[TMP1:%.*]] = load <2 x double>, ptr [[A:%.*]], align 8
30; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
31; CHECK:       for.body:
32; CHECK-NEXT:    [[I_020:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
33; CHECK-NEXT:    [[TMP2:%.*]] = phi <2 x double> [ [[TMP1]], [[ENTRY]] ], [ [[TMP5:%.*]], [[FOR_BODY]] ]
34; CHECK-NEXT:    [[TMP3:%.*]] = fadd <2 x double> [[TMP2]], splat (double 1.000000e+01)
35; CHECK-NEXT:    [[TMP4:%.*]] = fmul <2 x double> [[TMP3]], splat (double 4.000000e+00)
36; CHECK-NEXT:    [[TMP5]] = fadd <2 x double> [[TMP4]], splat (double 4.000000e+00)
37; CHECK-NEXT:    [[INC]] = add nsw i32 [[I_020]], 1
38; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[INC]], 100
39; CHECK-NEXT:    br i1 [[EXITCOND]], label [[FOR_END:%.*]], label [[FOR_BODY]]
40; CHECK:       for.end:
41; CHECK-NEXT:    store <2 x double> [[TMP5]], ptr [[B:%.*]], align 8
42; CHECK-NEXT:    [[TMP7:%.*]] = extractelement <2 x double> [[TMP4]], i32 0
43; CHECK-NEXT:    ret double [[TMP7]]
44;
45entry:
46  %arrayidx = getelementptr inbounds double, ptr %A, i64 1
47  %0 = load double, ptr %arrayidx, align 8
48  %1 = load double, ptr %A, align 8
49  br label %for.body
50
51for.body:                                         ; preds = %for.body, %entry
52  %i.020 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
53  %G.019 = phi double [ %1, %entry ], [ %add5, %for.body ]
54  %R.018 = phi double [ %0, %entry ], [ %add4, %for.body ]
55  %add = fadd double %R.018, 1.000000e+01
56  %add2 = fadd double %G.019, 1.000000e+01
57  %mul = fmul double %add, 4.000000e+00
58  %mul3 = fmul double %add2, 4.000000e+00
59  %add4 = fadd double %mul, 4.000000e+00
60  %add5 = fadd double %mul3, 4.000000e+00
61  %inc = add nsw i32 %i.020, 1
62  %exitcond = icmp eq i32 %inc, 100
63  br i1 %exitcond, label %for.end, label %for.body
64
65for.end:                                          ; preds = %for.body
66  store double %add5, ptr %B, align 8
67  %arrayidx7 = getelementptr inbounds double, ptr %B, i64 1
68  store double %add4, ptr %arrayidx7, align 8
69  ret double %mul3
70}
71
72; A need-to-gather entry cannot be an external use of the scalar element.
73; Instead the insertelement instructions of the need-to-gather entry are the
74; external users.
75; This test would assert because we would keep the scalar fpext and fadd alive.
76; PR18129
77
78define i32 @needtogather(ptr noalias %a, ptr noalias %b,  ptr noalias %c,
79; CHECK-LABEL: @needtogather(
80; CHECK-NEXT:  entry:
81; CHECK-NEXT:    [[TMP0:%.*]] = load i32, ptr [[D:%.*]], align 4
82; CHECK-NEXT:    [[CONV:%.*]] = sitofp i32 [[TMP0]] to float
83; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr [[C:%.*]], align 4
84; CHECK-NEXT:    [[SUB:%.*]] = fsub float 0.000000e+00, [[TMP1]]
85; CHECK-NEXT:    [[MUL:%.*]] = fmul float [[SUB]], 0.000000e+00
86; CHECK-NEXT:    [[ADD:%.*]] = fadd float [[CONV]], [[MUL]]
87; CHECK-NEXT:    [[CONV1:%.*]] = fpext float [[ADD]] to double
88; CHECK-NEXT:    [[SUB3:%.*]] = fsub float 1.000000e+00, [[TMP1]]
89; CHECK-NEXT:    [[MUL4:%.*]] = fmul float [[SUB3]], 0.000000e+00
90; CHECK-NEXT:    [[ADD5:%.*]] = fadd float [[CONV]], [[MUL4]]
91; CHECK-NEXT:    [[CONV6:%.*]] = fpext float [[ADD5]] to double
92; CHECK-NEXT:    [[TOBOOL:%.*]] = fcmp une float [[ADD]], 0.000000e+00
93; CHECK-NEXT:    br i1 [[TOBOOL]], label [[IF_THEN:%.*]], label [[IF_END:%.*]]
94; CHECK:       if.then:
95; CHECK-NEXT:    br label [[IF_END]]
96; CHECK:       if.end:
97; CHECK-NEXT:    [[STOREMERGE:%.*]] = phi double [ [[CONV6]], [[IF_THEN]] ], [ [[CONV1]], [[ENTRY:%.*]] ]
98; CHECK-NEXT:    [[E_0:%.*]] = phi double [ [[CONV1]], [[IF_THEN]] ], [ [[CONV6]], [[ENTRY]] ]
99; CHECK-NEXT:    store double [[STOREMERGE]], ptr [[A:%.*]], align 8
100; CHECK-NEXT:    [[CONV7:%.*]] = fptosi double [[E_0]] to i32
101; CHECK-NEXT:    store i32 [[CONV7]], ptr [[B:%.*]], align 4
102; CHECK-NEXT:    ret i32 undef
103;
104  ptr noalias %d) {
105entry:
106  %0 = load i32, ptr %d, align 4
107  %conv = sitofp i32 %0 to float
108  %1 = load float, ptr %c
109  %sub = fsub float 0.000000e+00, %1
110  %mul = fmul float %sub, 0.000000e+00
111  %add = fadd float %conv, %mul
112  %conv1 = fpext float %add to double
113  %sub3 = fsub float 1.000000e+00, %1
114  %mul4 = fmul float %sub3, 0.000000e+00
115  %add5 = fadd float %conv, %mul4
116  %conv6 = fpext float %add5 to double
117  %tobool = fcmp une float %add, 0.000000e+00
118  br i1 %tobool, label %if.then, label %if.end
119
120if.then:
121  br label %if.end
122
123if.end:
124  %storemerge = phi double [ %conv6, %if.then ], [ %conv1, %entry ]
125  %e.0 = phi double [ %conv1, %if.then ], [ %conv6, %entry ]
126  store double %storemerge, ptr %a, align 8
127  %conv7 = fptosi double %e.0 to i32
128  store i32 %conv7, ptr %b, align 4
129  ret i32 undef
130}
131