1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z10 | FileCheck %s
3; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z13 | FileCheck %s
4
5;
6; Test strict 64-bit square root.
7;
8
9declare double @llvm.experimental.constrained.sqrt.f64(double, metadata, metadata)
10
11; Check register square root.
12define double @f1(double %val) #0 {
13; CHECK-LABEL: f1:
14; CHECK:       # %bb.0:
15; CHECK-NEXT:    sqdbr %f0, %f0
16; CHECK-NEXT:    br %r14
17  %res = call double @llvm.experimental.constrained.sqrt.f64(
18                        double %val,
19                        metadata !"round.dynamic",
20                        metadata !"fpexcept.strict") #0
21  ret double %res
22}
23
24; Check the low end of the SQDB range.
25define double @f2(double *%ptr) #0 {
26; CHECK-LABEL: f2:
27; CHECK:       # %bb.0:
28; CHECK-NEXT:    sqdb %f0, 0(%r2)
29; CHECK-NEXT:    br %r14
30  %val = load double, double *%ptr
31  %res = call double @llvm.experimental.constrained.sqrt.f64(
32                        double %val,
33                        metadata !"round.dynamic",
34                        metadata !"fpexcept.strict") #0
35  ret double %res
36}
37
38; Check the high end of the aligned SQDB range.
39define double @f3(double *%base) #0 {
40; CHECK-LABEL: f3:
41; CHECK:       # %bb.0:
42; CHECK-NEXT:    sqdb %f0, 4088(%r2)
43; CHECK-NEXT:    br %r14
44  %ptr = getelementptr double, double *%base, i64 511
45  %val = load double, double *%ptr
46  %res = call double @llvm.experimental.constrained.sqrt.f64(
47                        double %val,
48                        metadata !"round.dynamic",
49                        metadata !"fpexcept.strict") #0
50  ret double %res
51}
52
53; Check the next doubleword up, which needs separate address logic.
54; Other sequences besides this one would be OK.
55define double @f4(double *%base) #0 {
56; CHECK-LABEL: f4:
57; CHECK:       # %bb.0:
58; CHECK-NEXT:    aghi %r2, 4096
59; CHECK-NEXT:    sqdb %f0, 0(%r2)
60; CHECK-NEXT:    br %r14
61  %ptr = getelementptr double, double *%base, i64 512
62  %val = load double, double *%ptr
63  %res = call double @llvm.experimental.constrained.sqrt.f64(
64                        double %val,
65                        metadata !"round.dynamic",
66                        metadata !"fpexcept.strict") #0
67  ret double %res
68}
69
70; Check negative displacements, which also need separate address logic.
71define double @f5(double *%base) #0 {
72; CHECK-LABEL: f5:
73; CHECK:       # %bb.0:
74; CHECK-NEXT:    aghi %r2, -8
75; CHECK-NEXT:    sqdb %f0, 0(%r2)
76; CHECK-NEXT:    br %r14
77  %ptr = getelementptr double, double *%base, i64 -1
78  %val = load double, double *%ptr
79  %res = call double @llvm.experimental.constrained.sqrt.f64(
80                        double %val,
81                        metadata !"round.dynamic",
82                        metadata !"fpexcept.strict") #0
83  ret double %res
84}
85
86; Check that SQDB allows indices.
87define double @f6(double *%base, i64 %index) #0 {
88; CHECK-LABEL: f6:
89; CHECK:       # %bb.0:
90; CHECK-NEXT:    sllg %r1, %r3, 3
91; CHECK-NEXT:    sqdb %f0, 800(%r1,%r2)
92; CHECK-NEXT:    br %r14
93  %ptr1 = getelementptr double, double *%base, i64 %index
94  %ptr2 = getelementptr double, double *%ptr1, i64 100
95  %val = load double, double *%ptr2
96  %res = call double @llvm.experimental.constrained.sqrt.f64(
97                        double %val,
98                        metadata !"round.dynamic",
99                        metadata !"fpexcept.strict") #0
100  ret double %res
101}
102
103attributes #0 = { strictfp }
104