1; Test load/store pairs that act as memcpys.
2;
3; RUN: llc < %s -mtriple=s390x-linux-gnu | FileCheck %s
4
5@g1src = global i8 1
6@g1dst = global i8 1
7@g2src = global i16 2
8@g2dst = global i16 2
9@g3 = global i32 3
10@g4 = global i64 4
11@g5src = external global fp128, align 16
12@g5dst = external global fp128, align 16
13
14; Test the simple i8 case.
15define void @f1(i8 *%ptr1) {
16; CHECK-LABEL: f1:
17; CHECK: mvc 1(1,%r2), 0(%r2)
18; CHECK: br %r14
19  %ptr2 = getelementptr i8, i8 *%ptr1, i64 1
20  %val = load i8, i8 *%ptr1
21  store i8 %val, i8 *%ptr2
22  ret void
23}
24
25; Test i8 cases where the value is zero-extended to 32 bits.
26define void @f2(i8 *%ptr1) {
27; CHECK-LABEL: f2:
28; CHECK: mvc 1(1,%r2), 0(%r2)
29; CHECK: br %r14
30  %ptr2 = getelementptr i8, i8 *%ptr1, i64 1
31  %val = load i8, i8 *%ptr1
32  %ext = zext i8 %val to i32
33  %trunc = trunc i32 %ext to i8
34  store i8 %trunc, i8 *%ptr2
35  ret void
36}
37
38; Test i8 cases where the value is zero-extended to 64 bits.
39define void @f3(i8 *%ptr1) {
40; CHECK-LABEL: f3:
41; CHECK: mvc 1(1,%r2), 0(%r2)
42; CHECK: br %r14
43  %ptr2 = getelementptr i8, i8 *%ptr1, i64 1
44  %val = load i8, i8 *%ptr1
45  %ext = zext i8 %val to i64
46  %trunc = trunc i64 %ext to i8
47  store i8 %trunc, i8 *%ptr2
48  ret void
49}
50
51; Test i8 cases where the value is sign-extended to 32 bits.
52define void @f4(i8 *%ptr1) {
53; CHECK-LABEL: f4:
54; CHECK: mvc 1(1,%r2), 0(%r2)
55; CHECK: br %r14
56  %ptr2 = getelementptr i8, i8 *%ptr1, i64 1
57  %val = load i8, i8 *%ptr1
58  %ext = sext i8 %val to i32
59  %trunc = trunc i32 %ext to i8
60  store i8 %trunc, i8 *%ptr2
61  ret void
62}
63
64; Test i8 cases where the value is sign-extended to 64 bits.
65define void @f5(i8 *%ptr1) {
66; CHECK-LABEL: f5:
67; CHECK: mvc 1(1,%r2), 0(%r2)
68; CHECK: br %r14
69  %ptr2 = getelementptr i8, i8 *%ptr1, i64 1
70  %val = load i8, i8 *%ptr1
71  %ext = sext i8 %val to i64
72  %trunc = trunc i64 %ext to i8
73  store i8 %trunc, i8 *%ptr2
74  ret void
75}
76
77; Test the simple i16 case.
78define void @f6(i16 *%ptr1) {
79; CHECK-LABEL: f6:
80; CHECK: mvc 2(2,%r2), 0(%r2)
81; CHECK: br %r14
82  %ptr2 = getelementptr i16, i16 *%ptr1, i64 1
83  %val = load i16, i16 *%ptr1
84  store i16 %val, i16 *%ptr2
85  ret void
86}
87
88; Test i16 cases where the value is zero-extended to 32 bits.
89define void @f7(i16 *%ptr1) {
90; CHECK-LABEL: f7:
91; CHECK: mvc 2(2,%r2), 0(%r2)
92; CHECK: br %r14
93  %ptr2 = getelementptr i16, i16 *%ptr1, i64 1
94  %val = load i16, i16 *%ptr1
95  %ext = zext i16 %val to i32
96  %trunc = trunc i32 %ext to i16
97  store i16 %trunc, i16 *%ptr2
98  ret void
99}
100
101; Test i16 cases where the value is zero-extended to 64 bits.
102define void @f8(i16 *%ptr1) {
103; CHECK-LABEL: f8:
104; CHECK: mvc 2(2,%r2), 0(%r2)
105; CHECK: br %r14
106  %ptr2 = getelementptr i16, i16 *%ptr1, i64 1
107  %val = load i16, i16 *%ptr1
108  %ext = zext i16 %val to i64
109  %trunc = trunc i64 %ext to i16
110  store i16 %trunc, i16 *%ptr2
111  ret void
112}
113
114; Test i16 cases where the value is sign-extended to 32 bits.
115define void @f9(i16 *%ptr1) {
116; CHECK-LABEL: f9:
117; CHECK: mvc 2(2,%r2), 0(%r2)
118; CHECK: br %r14
119  %ptr2 = getelementptr i16, i16 *%ptr1, i64 1
120  %val = load i16, i16 *%ptr1
121  %ext = sext i16 %val to i32
122  %trunc = trunc i32 %ext to i16
123  store i16 %trunc, i16 *%ptr2
124  ret void
125}
126
127; Test i16 cases where the value is sign-extended to 64 bits.
128define void @f10(i16 *%ptr1) {
129; CHECK-LABEL: f10:
130; CHECK: mvc 2(2,%r2), 0(%r2)
131; CHECK: br %r14
132  %ptr2 = getelementptr i16, i16 *%ptr1, i64 1
133  %val = load i16, i16 *%ptr1
134  %ext = sext i16 %val to i64
135  %trunc = trunc i64 %ext to i16
136  store i16 %trunc, i16 *%ptr2
137  ret void
138}
139
140; Test the simple i32 case.
141define void @f11(i32 *%ptr1) {
142; CHECK-LABEL: f11:
143; CHECK: mvc 4(4,%r2), 0(%r2)
144; CHECK: br %r14
145  %ptr2 = getelementptr i32, i32 *%ptr1, i64 1
146  %val = load i32, i32 *%ptr1
147  store i32 %val, i32 *%ptr2
148  ret void
149}
150
151; Test i32 cases where the value is zero-extended to 64 bits.
152define void @f12(i32 *%ptr1) {
153; CHECK-LABEL: f12:
154; CHECK: mvc 4(4,%r2), 0(%r2)
155; CHECK: br %r14
156  %ptr2 = getelementptr i32, i32 *%ptr1, i64 1
157  %val = load i32, i32 *%ptr1
158  %ext = zext i32 %val to i64
159  %trunc = trunc i64 %ext to i32
160  store i32 %trunc, i32 *%ptr2
161  ret void
162}
163
164; Test i32 cases where the value is sign-extended to 64 bits.
165define void @f13(i32 *%ptr1) {
166; CHECK-LABEL: f13:
167; CHECK: mvc 4(4,%r2), 0(%r2)
168; CHECK: br %r14
169  %ptr2 = getelementptr i32, i32 *%ptr1, i64 1
170  %val = load i32, i32 *%ptr1
171  %ext = sext i32 %val to i64
172  %trunc = trunc i64 %ext to i32
173  store i32 %trunc, i32 *%ptr2
174  ret void
175}
176
177; Test the i64 case.
178define void @f14(i64 *%ptr1) {
179; CHECK-LABEL: f14:
180; CHECK: mvc 8(8,%r2), 0(%r2)
181; CHECK: br %r14
182  %ptr2 = getelementptr i64, i64 *%ptr1, i64 1
183  %val = load i64, i64 *%ptr1
184  store i64 %val, i64 *%ptr2
185  ret void
186}
187
188; Test the f32 case.
189define void @f15(float *%ptr1) {
190; CHECK-LABEL: f15:
191; CHECK: mvc 4(4,%r2), 0(%r2)
192; CHECK: br %r14
193  %ptr2 = getelementptr float, float *%ptr1, i64 1
194  %val = load float, float *%ptr1
195  store float %val, float *%ptr2
196  ret void
197}
198
199; Test the f64 case.
200define void @f16(double *%ptr1) {
201; CHECK-LABEL: f16:
202; CHECK: mvc 8(8,%r2), 0(%r2)
203; CHECK: br %r14
204  %ptr2 = getelementptr double, double *%ptr1, i64 1
205  %val = load double, double *%ptr1
206  store double %val, double *%ptr2
207  ret void
208}
209
210; Test the f128 case.
211define void @f17(fp128 *%ptr1) {
212; CHECK-LABEL: f17:
213; CHECK: mvc 16(16,%r2), 0(%r2)
214; CHECK: br %r14
215  %ptr2 = getelementptr fp128, fp128 *%ptr1, i64 1
216  %val = load fp128, fp128 *%ptr1
217  store fp128 %val, fp128 *%ptr2
218  ret void
219}
220
221; Make sure that we don't use MVC if the load is volatile.
222define void @f18(i64 *%ptr1) {
223; CHECK-LABEL: f18:
224; CHECK-NOT: mvc
225; CHECK: br %r14
226  %ptr2 = getelementptr i64, i64 *%ptr1, i64 1
227  %val = load volatile i64, i64 *%ptr1
228  store i64 %val, i64 *%ptr2
229  ret void
230}
231
232; ...likewise the store.
233define void @f19(i64 *%ptr1) {
234; CHECK-LABEL: f19:
235; CHECK-NOT: mvc
236; CHECK: br %r14
237  %ptr2 = getelementptr i64, i64 *%ptr1, i64 1
238  %val = load i64, i64 *%ptr1
239  store volatile i64 %val, i64 *%ptr2
240  ret void
241}
242
243; Test that MVC is not used for aligned loads and stores if there is
244; no way of telling whether they alias.  We don't want to use MVC in
245; cases where the addresses could be equal.
246define void @f20(i64 *%ptr1, i64 *%ptr2) {
247; CHECK-LABEL: f20:
248; CHECK-NOT: mvc
249; CHECK: br %r14
250  %val = load i64, i64 *%ptr1
251  store i64 %val, i64 *%ptr2
252  ret void
253}
254
255; ...and again for unaligned loads and stores.
256define void @f21(i64 *%ptr1, i64 *%ptr2) {
257; CHECK-LABEL: f21:
258; CHECK-NOT: mvc
259; CHECK: br %r14
260  %val = load i64, i64 *%ptr1, align 2
261  store i64 %val, i64 *%ptr2, align 2
262  ret void
263}
264
265; Test a case where there is definite overlap.
266define void @f22(i64 %base) {
267; CHECK-LABEL: f22:
268; CHECK-NOT: mvc
269; CHECK: br %r14
270  %add = add i64 %base, 1
271  %ptr1 = inttoptr i64 %base to i64 *
272  %ptr2 = inttoptr i64 %add to i64 *
273  %val = load i64, i64 *%ptr1, align 1
274  store i64 %val, i64 *%ptr2, align 1
275  ret void
276}
277
278; Test that we can use MVC for global addresses for i8.
279define void @f23(i8 *%ptr) {
280; CHECK-LABEL: f23:
281; CHECK-DAG: larl [[SRC:%r[0-5]]], g1src
282; CHECK-DAG: larl [[DST:%r[0-5]]], g1dst
283; CHECK: mvc 0(1,[[DST]]), 0([[SRC]])
284; CHECK: br %r14
285  %val = load i8, i8 *@g1src
286  store i8 %val, i8 *@g1dst
287  ret void
288}
289
290; Test that we use LHRL and STHRL for i16.
291define void @f24(i16 *%ptr) {
292; CHECK-LABEL: f24:
293; CHECK: lhrl [[REG:%r[0-5]]], g2src
294; CHECK: sthrl [[REG]], g2dst
295; CHECK: br %r14
296  %val = load i16, i16 *@g2src
297  store i16 %val, i16 *@g2dst
298  ret void
299}
300
301; Test that we use LRL for i32.
302define void @f25(i32 *%ptr) {
303; CHECK-LABEL: f25:
304; CHECK: lrl [[REG:%r[0-5]]], g3
305; CHECK: st [[REG]], 0(%r2)
306; CHECK: br %r14
307  %val = load i32, i32 *@g3
308  store i32 %val, i32 *%ptr
309  ret void
310}
311
312; ...likewise STRL.
313define void @f26(i32 *%ptr) {
314; CHECK-LABEL: f26:
315; CHECK: l [[REG:%r[0-5]]], 0(%r2)
316; CHECK: strl [[REG]], g3
317; CHECK: br %r14
318  %val = load i32, i32 *%ptr
319  store i32 %val, i32 *@g3
320  ret void
321}
322
323; Test that we use LGRL for i64.
324define void @f27(i64 *%ptr) {
325; CHECK-LABEL: f27:
326; CHECK: lgrl [[REG:%r[0-5]]], g4
327; CHECK: stg [[REG]], 0(%r2)
328; CHECK: br %r14
329  %val = load i64, i64 *@g4
330  store i64 %val, i64 *%ptr
331  ret void
332}
333
334; ...likewise STGRL.
335define void @f28(i64 *%ptr) {
336; CHECK-LABEL: f28:
337; CHECK: lg [[REG:%r[0-5]]], 0(%r2)
338; CHECK: stgrl [[REG]], g4
339; CHECK: br %r14
340  %val = load i64, i64 *%ptr
341  store i64 %val, i64 *@g4
342  ret void
343}
344
345; Test that we can use MVC for global addresses for fp128.
346define void @f29(fp128 *%ptr) {
347; CHECK-LABEL: f29:
348; CHECK-DAG: larl [[SRC:%r[0-5]]], g5src
349; CHECK-DAG: larl [[DST:%r[0-5]]], g5dst
350; CHECK: mvc 0(16,[[DST]]), 0([[SRC]])
351; CHECK: br %r14
352  %val = load fp128, fp128 *@g5src, align 16
353  store fp128 %val, fp128 *@g5dst, align 16
354  ret void
355}
356
357; Test a case where offset disambiguation is enough.
358define void @f30(i64 *%ptr1) {
359; CHECK-LABEL: f30:
360; CHECK: mvc 8(8,%r2), 0(%r2)
361; CHECK: br %r14
362  %ptr2 = getelementptr i64, i64 *%ptr1, i64 1
363  %val = load i64, i64 *%ptr1, align 1
364  store i64 %val, i64 *%ptr2, align 1
365  ret void
366}
367
368; Test f21 in cases where TBAA tells us there is no alias.
369define void @f31(i64 *%ptr1, i64 *%ptr2) {
370; CHECK-LABEL: f31:
371; CHECK: mvc 0(8,%r3), 0(%r2)
372; CHECK: br %r14
373  %val = load i64, i64 *%ptr1, align 2, !tbaa !1
374  store i64 %val, i64 *%ptr2, align 2, !tbaa !2
375  ret void
376}
377
378; Test f21 in cases where TBAA is present but doesn't help.
379define void @f32(i64 *%ptr1, i64 *%ptr2) {
380; CHECK-LABEL: f32:
381; CHECK-NOT: mvc
382; CHECK: br %r14
383  %val = load i64, i64 *%ptr1, align 2, !tbaa !1
384  store i64 %val, i64 *%ptr2, align 2, !tbaa !1
385  ret void
386}
387
388!0 = !{ !"root" }
389!1 = !{ !3, !3, i64 0 }
390!2 = !{ !4, !4, i64 0 }
391!3 = !{ !"set1", !0 }
392!4 = !{ !"set2", !0 }
393