1; Test 32-bit shifts left.
2;
3; RUN: llc < %s -mtriple=s390x-linux-gnu | FileCheck %s
4
5; Check the low end of the SLLG range.
6define i64 @f1(i64 %a) {
7; CHECK-LABEL: f1:
8; CHECK: sllg %r2, %r2, 1
9; CHECK: br %r14
10  %shift = shl i64 %a, 1
11  ret i64 %shift
12}
13
14; Check the high end of the defined SLLG range.
15define i64 @f2(i64 %a) {
16; CHECK-LABEL: f2:
17; CHECK: sllg %r2, %r2, 63
18; CHECK: br %r14
19  %shift = shl i64 %a, 63
20  ret i64 %shift
21}
22
23; We don't generate shifts by out-of-range values.
24define i64 @f3(i64 %a) {
25; CHECK-LABEL: f3:
26; CHECK-NOT: sllg
27; CHECK: br %r14
28  %shift = shl i64 %a, 64
29  ret i64 %shift
30}
31
32; Check variable shifts.
33define i64 @f4(i64 %a, i64 %amt) {
34; CHECK-LABEL: f4:
35; CHECK: sllg %r2, %r2, 0(%r3)
36; CHECK: br %r14
37  %shift = shl i64 %a, %amt
38  ret i64 %shift
39}
40
41; Check shift amounts that have a constant term.
42define i64 @f5(i64 %a, i64 %amt) {
43; CHECK-LABEL: f5:
44; CHECK: sllg %r2, %r2, 10(%r3)
45; CHECK: br %r14
46  %add = add i64 %amt, 10
47  %shift = shl i64 %a, %add
48  ret i64 %shift
49}
50
51; ...and again with a sign-extended 32-bit shift amount.
52define i64 @f6(i64 %a, i32 %amt) {
53; CHECK-LABEL: f6:
54; CHECK: sllg %r2, %r2, 10(%r3)
55; CHECK: br %r14
56  %add = add i32 %amt, 10
57  %addext = sext i32 %add to i64
58  %shift = shl i64 %a, %addext
59  ret i64 %shift
60}
61
62; ...and now with a zero-extended 32-bit shift amount.
63define i64 @f7(i64 %a, i32 %amt) {
64; CHECK-LABEL: f7:
65; CHECK: sllg %r2, %r2, 10(%r3)
66; CHECK: br %r14
67  %add = add i32 %amt, 10
68  %addext = zext i32 %add to i64
69  %shift = shl i64 %a, %addext
70  ret i64 %shift
71}
72
73; Check shift amounts that have the largest in-range constant term.  We could
74; mask the amount instead.
75define i64 @f8(i64 %a, i64 %amt) {
76; CHECK-LABEL: f8:
77; CHECK: sllg %r2, %r2, 524287(%r3)
78; CHECK: br %r14
79  %add = add i64 %amt, 524287
80  %shift = shl i64 %a, %add
81  ret i64 %shift
82}
83
84; Check the next value up, which without masking must use a separate
85; addition.
86define i64 @f9(i64 %a, i64 %amt) {
87; CHECK-LABEL: f9:
88; CHECK: a{{g?}}fi %r3, 524288
89; CHECK: sllg %r2, %r2, 0(%r3)
90; CHECK: br %r14
91  %add = add i64 %amt, 524288
92  %shift = shl i64 %a, %add
93  ret i64 %shift
94}
95
96; Check cases where 1 is subtracted from the shift amount.
97define i64 @f10(i64 %a, i64 %amt) {
98; CHECK-LABEL: f10:
99; CHECK: sllg %r2, %r2, -1(%r3)
100; CHECK: br %r14
101  %sub = sub i64 %amt, 1
102  %shift = shl i64 %a, %sub
103  ret i64 %shift
104}
105
106; Check the lowest value that can be subtracted from the shift amount.
107; Again, we could mask the shift amount instead.
108define i64 @f11(i64 %a, i64 %amt) {
109; CHECK-LABEL: f11:
110; CHECK: sllg %r2, %r2, -524288(%r3)
111; CHECK: br %r14
112  %sub = sub i64 %amt, 524288
113  %shift = shl i64 %a, %sub
114  ret i64 %shift
115}
116
117; Check the next value down, which without masking must use a separate
118; addition.
119define i64 @f12(i64 %a, i64 %amt) {
120; CHECK-LABEL: f12:
121; CHECK: a{{g?}}fi %r3, -524289
122; CHECK: sllg %r2, %r2, 0(%r3)
123; CHECK: br %r14
124  %sub = sub i64 %amt, 524289
125  %shift = shl i64 %a, %sub
126  ret i64 %shift
127}
128
129; Check that we don't try to generate "indexed" shifts.
130define i64 @f13(i64 %a, i64 %b, i64 %c) {
131; CHECK-LABEL: f13:
132; CHECK: a{{g?}}r {{%r3, %r4|%r4, %r3}}
133; CHECK: sllg %r2, %r2, 0({{%r[34]}})
134; CHECK: br %r14
135  %add = add i64 %b, %c
136  %shift = shl i64 %a, %add
137  ret i64 %shift
138}
139
140; Check that the shift amount uses an address register.  It cannot be in %r0.
141define i64 @f14(i64 %a, i64 *%ptr) {
142; CHECK-LABEL: f14:
143; CHECK: l %r1, 4(%r3)
144; CHECK: sllg %r2, %r2, 0(%r1)
145; CHECK: br %r14
146  %amt = load i64 *%ptr
147  %shift = shl i64 %a, %amt
148  ret i64 %shift
149}
150