xref: /llvm-project/llvm/test/CodeGen/PowerPC/builtins-ppc-p10vsx.ll (revision a2b5117df75d1be8a65b2a86d5f75a59a8565fe6) 
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -verify-machineinstrs -mtriple=powerpc64le-unknown-linux-gnu \
3; RUN:   -mcpu=pwr10 -ppc-asm-full-reg-names -ppc-vsr-nums-as-vr < %s | \
4; RUN:   FileCheck %s --check-prefixes=CHECK,CHECK-LE
5; RUN: llc -verify-machineinstrs -mtriple=powerpc64-unknown-linux-gnu \
6; RUN:   -mcpu=pwr10 -ppc-asm-full-reg-names -ppc-vsr-nums-as-vr < %s | \
7; RUN:   FileCheck %s --check-prefixes=CHECK,CHECK-LINUXBE
8; RUN: llc -verify-machineinstrs -mtriple=powerpc64le-unknown-linux-gnu -O0 \
9; RUN:   -mcpu=pwr10 -ppc-asm-full-reg-names -ppc-vsr-nums-as-vr < %s | \
10; RUN:   FileCheck %s --check-prefixes=CHECK,CHECK-O0
11; RUN: llc -verify-machineinstrs -mtriple=powerpc64-ibm-aix-xcoff \
12; RUN:   -mcpu=pwr10 -ppc-asm-full-reg-names -ppc-vsr-nums-as-vr < %s | \
13; RUN:   FileCheck %s --check-prefixes=CHECK,CHECK-AIXBE
14
15; These test cases aims to test the builtins for the Power10 VSX vector
16; instructions introduced in ISA 3.1.
17
18declare i32 @llvm.ppc.vsx.xvtlsbb(<16 x i8>, i32)
19
20define signext i32 @test_vec_test_lsbb_all_ones(<16 x i8> %vuca) {
21; CHECK-LABEL: test_vec_test_lsbb_all_ones:
22; CHECK:       # %bb.0: # %entry
23; CHECK-NEXT:    xvtlsbb cr0, v2
24; CHECK-NEXT:    mfocrf r3, 128
25; CHECK-NEXT:    srwi r3, r3, 31
26; CHECK-NEXT:    extsw r3, r3
27; CHECK-NEXT:    blr
28entry:
29  %0 = tail call i32 @llvm.ppc.vsx.xvtlsbb(<16 x i8> %vuca, i32 1)
30  ret i32 %0
31}
32
33define signext i32 @test_vec_test_lsbb_all_zeros(<16 x i8> %vuca) {
34; CHECK-LABEL: test_vec_test_lsbb_all_zeros:
35; CHECK:       # %bb.0: # %entry
36; CHECK-NEXT:    xvtlsbb cr0, v2
37; CHECK-NEXT:    mfocrf r3, 128
38; CHECK-NEXT:    rlwinm r3, r3, 3, 31, 31
39; CHECK-NEXT:    extsw r3, r3
40; CHECK-NEXT:    blr
41entry:
42  %0 = tail call i32 @llvm.ppc.vsx.xvtlsbb(<16 x i8> %vuca, i32 0)
43  ret i32 %0
44}
45
46define void @vec_xst_trunc_sc(<1 x i128> %__vec, i64 %__offset, ptr nocapture %__ptr) {
47; CHECK-LE-LABEL: vec_xst_trunc_sc:
48; CHECK-LE:       # %bb.0: # %entry
49; CHECK-LE-NEXT:    stxvrbx v2, r6, r5
50; CHECK-LE-NEXT:    blr
51;
52; CHECK-LINUXBE-LABEL: vec_xst_trunc_sc:
53; CHECK-LINUXBE:       # %bb.0: # %entry
54; CHECK-LINUXBE-NEXT:    vsldoi v2, v2, v2, 9
55; CHECK-LINUXBE-NEXT:    stxsibx v2, r6, r5
56; CHECK-LINUXBE-NEXT:    blr
57;
58; CHECK-O0-LABEL: vec_xst_trunc_sc:
59; CHECK-O0:       # %bb.0: # %entry
60; CHECK-O0-NEXT:    li r3, 0
61; CHECK-O0-NEXT:    vextubrx r3, r3, v2
62; CHECK-O0-NEXT:    # kill: def $r3 killed $r3 killed $x3
63; CHECK-O0-NEXT:    add r4, r6, r5
64; CHECK-O0-NEXT:    stb r3, 0(r4)
65; CHECK-O0-NEXT:    blr
66;
67; CHECK-AIXBE-LABEL: vec_xst_trunc_sc:
68; CHECK-AIXBE:       # %bb.0: # %entry
69; CHECK-AIXBE-NEXT:    vsldoi v2, v2, v2, 9
70; CHECK-AIXBE-NEXT:    stxsibx v2, r4, r3
71; CHECK-AIXBE-NEXT:    blr
72entry:
73  %0 = bitcast <1 x i128> %__vec to <16 x i8>
74  %conv = extractelement <16 x i8> %0, i32 0
75  %add.ptr = getelementptr inbounds i8, ptr %__ptr, i64 %__offset
76  store i8 %conv, ptr %add.ptr, align 1
77  ret void
78}
79
80define void @vec_xst_trunc_uc(<1 x i128> %__vec, i64 %__offset, ptr nocapture %__ptr) {
81; CHECK-LE-LABEL: vec_xst_trunc_uc:
82; CHECK-LE:       # %bb.0: # %entry
83; CHECK-LE-NEXT:    stxvrbx v2, r6, r5
84; CHECK-LE-NEXT:    blr
85;
86; CHECK-LINUXBE-LABEL: vec_xst_trunc_uc:
87; CHECK-LINUXBE:       # %bb.0: # %entry
88; CHECK-LINUXBE-NEXT:    vsldoi v2, v2, v2, 9
89; CHECK-LINUXBE-NEXT:    stxsibx v2, r6, r5
90; CHECK-LINUXBE-NEXT:    blr
91;
92; CHECK-O0-LABEL: vec_xst_trunc_uc:
93; CHECK-O0:       # %bb.0: # %entry
94; CHECK-O0-NEXT:    li r3, 0
95; CHECK-O0-NEXT:    vextubrx r3, r3, v2
96; CHECK-O0-NEXT:    # kill: def $r3 killed $r3 killed $x3
97; CHECK-O0-NEXT:    add r4, r6, r5
98; CHECK-O0-NEXT:    stb r3, 0(r4)
99; CHECK-O0-NEXT:    blr
100;
101; CHECK-AIXBE-LABEL: vec_xst_trunc_uc:
102; CHECK-AIXBE:       # %bb.0: # %entry
103; CHECK-AIXBE-NEXT:    vsldoi v2, v2, v2, 9
104; CHECK-AIXBE-NEXT:    stxsibx v2, r4, r3
105; CHECK-AIXBE-NEXT:    blr
106entry:
107  %0 = bitcast <1 x i128> %__vec to <16 x i8>
108  %conv = extractelement <16 x i8> %0, i32 0
109  %add.ptr = getelementptr inbounds i8, ptr %__ptr, i64 %__offset
110  store i8 %conv, ptr %add.ptr, align 1
111  ret void
112}
113
114define void @vec_xst_trunc_ss(<1 x i128> %__vec, i64 %__offset, ptr nocapture %__ptr) {
115; CHECK-LE-LABEL: vec_xst_trunc_ss:
116; CHECK-LE:       # %bb.0: # %entry
117; CHECK-LE-NEXT:    sldi r3, r5, 1
118; CHECK-LE-NEXT:    stxvrhx v2, r6, r3
119; CHECK-LE-NEXT:    blr
120;
121; CHECK-LINUXBE-LABEL: vec_xst_trunc_ss:
122; CHECK-LINUXBE:       # %bb.0: # %entry
123; CHECK-LINUXBE-NEXT:    vsldoi v2, v2, v2, 10
124; CHECK-LINUXBE-NEXT:    sldi r3, r5, 1
125; CHECK-LINUXBE-NEXT:    stxsihx v2, r6, r3
126; CHECK-LINUXBE-NEXT:    blr
127;
128; CHECK-O0-LABEL: vec_xst_trunc_ss:
129; CHECK-O0:       # %bb.0: # %entry
130; CHECK-O0-NEXT:    li r3, 0
131; CHECK-O0-NEXT:    vextuhrx r3, r3, v2
132; CHECK-O0-NEXT:    # kill: def $r3 killed $r3 killed $x3
133; CHECK-O0-NEXT:    sldi r4, r5, 1
134; CHECK-O0-NEXT:    add r4, r6, r4
135; CHECK-O0-NEXT:    sth r3, 0(r4)
136; CHECK-O0-NEXT:    blr
137;
138; CHECK-AIXBE-LABEL: vec_xst_trunc_ss:
139; CHECK-AIXBE:       # %bb.0: # %entry
140; CHECK-AIXBE-NEXT:    vsldoi v2, v2, v2, 10
141; CHECK-AIXBE-NEXT:    sldi r3, r3, 1
142; CHECK-AIXBE-NEXT:    stxsihx v2, r4, r3
143; CHECK-AIXBE-NEXT:    blr
144entry:
145  %0 = bitcast <1 x i128> %__vec to <8 x i16>
146  %conv = extractelement <8 x i16> %0, i32 0
147  %add.ptr = getelementptr inbounds i16, ptr %__ptr, i64 %__offset
148  store i16 %conv, ptr %add.ptr, align 2
149  ret void
150}
151
152define void @vec_xst_trunc_us(<1 x i128> %__vec, i64 %__offset, ptr nocapture %__ptr) {
153; CHECK-LE-LABEL: vec_xst_trunc_us:
154; CHECK-LE:       # %bb.0: # %entry
155; CHECK-LE-NEXT:    sldi r3, r5, 1
156; CHECK-LE-NEXT:    stxvrhx v2, r6, r3
157; CHECK-LE-NEXT:    blr
158;
159; CHECK-LINUXBE-LABEL: vec_xst_trunc_us:
160; CHECK-LINUXBE:       # %bb.0: # %entry
161; CHECK-LINUXBE-NEXT:    vsldoi v2, v2, v2, 10
162; CHECK-LINUXBE-NEXT:    sldi r3, r5, 1
163; CHECK-LINUXBE-NEXT:    stxsihx v2, r6, r3
164; CHECK-LINUXBE-NEXT:    blr
165;
166; CHECK-O0-LABEL: vec_xst_trunc_us:
167; CHECK-O0:       # %bb.0: # %entry
168; CHECK-O0-NEXT:    li r3, 0
169; CHECK-O0-NEXT:    vextuhrx r3, r3, v2
170; CHECK-O0-NEXT:    # kill: def $r3 killed $r3 killed $x3
171; CHECK-O0-NEXT:    sldi r4, r5, 1
172; CHECK-O0-NEXT:    add r4, r6, r4
173; CHECK-O0-NEXT:    sth r3, 0(r4)
174; CHECK-O0-NEXT:    blr
175;
176; CHECK-AIXBE-LABEL: vec_xst_trunc_us:
177; CHECK-AIXBE:       # %bb.0: # %entry
178; CHECK-AIXBE-NEXT:    vsldoi v2, v2, v2, 10
179; CHECK-AIXBE-NEXT:    sldi r3, r3, 1
180; CHECK-AIXBE-NEXT:    stxsihx v2, r4, r3
181; CHECK-AIXBE-NEXT:    blr
182entry:
183  %0 = bitcast <1 x i128> %__vec to <8 x i16>
184  %conv = extractelement <8 x i16> %0, i32 0
185  %add.ptr = getelementptr inbounds i16, ptr %__ptr, i64 %__offset
186  store i16 %conv, ptr %add.ptr, align 2
187  ret void
188}
189
190define void @vec_xst_trunc_si(<1 x i128> %__vec, i64 %__offset, ptr nocapture %__ptr) {
191; CHECK-LE-LABEL: vec_xst_trunc_si:
192; CHECK-LE:       # %bb.0: # %entry
193; CHECK-LE-NEXT:    sldi r3, r5, 2
194; CHECK-LE-NEXT:    stxvrwx v2, r6, r3
195; CHECK-LE-NEXT:    blr
196;
197; CHECK-LINUXBE-LABEL: vec_xst_trunc_si:
198; CHECK-LINUXBE:       # %bb.0: # %entry
199; CHECK-LINUXBE-NEXT:    xxsldwi vs0, v2, v2, 3
200; CHECK-LINUXBE-NEXT:    sldi r3, r5, 2
201; CHECK-LINUXBE-NEXT:    stfiwx f0, r6, r3
202; CHECK-LINUXBE-NEXT:    blr
203;
204; CHECK-O0-LABEL: vec_xst_trunc_si:
205; CHECK-O0:       # %bb.0: # %entry
206; CHECK-O0-NEXT:    li r3, 0
207; CHECK-O0-NEXT:    vextuwrx r3, r3, v2
208; CHECK-O0-NEXT:    # kill: def $r3 killed $r3 killed $x3
209; CHECK-O0-NEXT:    sldi r4, r5, 2
210; CHECK-O0-NEXT:    add r4, r6, r4
211; CHECK-O0-NEXT:    stw r3, 0(r4)
212; CHECK-O0-NEXT:    blr
213;
214; CHECK-AIXBE-LABEL: vec_xst_trunc_si:
215; CHECK-AIXBE:       # %bb.0: # %entry
216; CHECK-AIXBE-NEXT:    xxsldwi vs0, v2, v2, 3
217; CHECK-AIXBE-NEXT:    sldi r3, r3, 2
218; CHECK-AIXBE-NEXT:    stfiwx f0, r4, r3
219; CHECK-AIXBE-NEXT:    blr
220entry:
221  %0 = bitcast <1 x i128> %__vec to <4 x i32>
222  %conv = extractelement <4 x i32> %0, i32 0
223  %add.ptr = getelementptr inbounds i32, ptr %__ptr, i64 %__offset
224  store i32 %conv, ptr %add.ptr, align 4
225  ret void
226}
227
228define void @vec_xst_trunc_ui(<1 x i128> %__vec, i64 %__offset, ptr nocapture %__ptr) {
229; CHECK-LE-LABEL: vec_xst_trunc_ui:
230; CHECK-LE:       # %bb.0: # %entry
231; CHECK-LE-NEXT:    sldi r3, r5, 2
232; CHECK-LE-NEXT:    stxvrwx v2, r6, r3
233; CHECK-LE-NEXT:    blr
234;
235; CHECK-LINUXBE-LABEL: vec_xst_trunc_ui:
236; CHECK-LINUXBE:       # %bb.0: # %entry
237; CHECK-LINUXBE-NEXT:    xxsldwi vs0, v2, v2, 3
238; CHECK-LINUXBE-NEXT:    sldi r3, r5, 2
239; CHECK-LINUXBE-NEXT:    stfiwx f0, r6, r3
240; CHECK-LINUXBE-NEXT:    blr
241;
242; CHECK-O0-LABEL: vec_xst_trunc_ui:
243; CHECK-O0:       # %bb.0: # %entry
244; CHECK-O0-NEXT:    li r3, 0
245; CHECK-O0-NEXT:    vextuwrx r3, r3, v2
246; CHECK-O0-NEXT:    # kill: def $r3 killed $r3 killed $x3
247; CHECK-O0-NEXT:    sldi r4, r5, 2
248; CHECK-O0-NEXT:    add r4, r6, r4
249; CHECK-O0-NEXT:    stw r3, 0(r4)
250; CHECK-O0-NEXT:    blr
251;
252; CHECK-AIXBE-LABEL: vec_xst_trunc_ui:
253; CHECK-AIXBE:       # %bb.0: # %entry
254; CHECK-AIXBE-NEXT:    xxsldwi vs0, v2, v2, 3
255; CHECK-AIXBE-NEXT:    sldi r3, r3, 2
256; CHECK-AIXBE-NEXT:    stfiwx f0, r4, r3
257; CHECK-AIXBE-NEXT:    blr
258entry:
259  %0 = bitcast <1 x i128> %__vec to <4 x i32>
260  %conv = extractelement <4 x i32> %0, i32 0
261  %add.ptr = getelementptr inbounds i32, ptr %__ptr, i64 %__offset
262  store i32 %conv, ptr %add.ptr, align 4
263  ret void
264}
265
266define void @vec_xst_trunc_sll(<1 x i128> %__vec, i64 %__offset, ptr nocapture %__ptr)  {
267; CHECK-LE-LABEL: vec_xst_trunc_sll:
268; CHECK-LE:       # %bb.0: # %entry
269; CHECK-LE-NEXT:    sldi r3, r5, 3
270; CHECK-LE-NEXT:    stxvrdx v2, r6, r3
271; CHECK-LE-NEXT:    blr
272;
273; CHECK-LINUXBE-LABEL: vec_xst_trunc_sll:
274; CHECK-LINUXBE:       # %bb.0: # %entry
275; CHECK-LINUXBE-NEXT:    sldi r3, r5, 3
276; CHECK-LINUXBE-NEXT:    stxsdx v2, r6, r3
277; CHECK-LINUXBE-NEXT:    blr
278;
279; CHECK-O0-LABEL: vec_xst_trunc_sll:
280; CHECK-O0:       # %bb.0: # %entry
281; CHECK-O0-NEXT:    mfvsrld r3, v2
282; CHECK-O0-NEXT:    sldi r4, r5, 3
283; CHECK-O0-NEXT:    add r4, r6, r4
284; CHECK-O0-NEXT:    std r3, 0(r4)
285; CHECK-O0-NEXT:    blr
286;
287; CHECK-AIXBE-LABEL: vec_xst_trunc_sll:
288; CHECK-AIXBE:       # %bb.0: # %entry
289; CHECK-AIXBE-NEXT:    sldi r3, r3, 3
290; CHECK-AIXBE-NEXT:    stxsdx v2, r4, r3
291; CHECK-AIXBE-NEXT:    blr
292entry:
293  %0 = bitcast <1 x i128> %__vec to <2 x i64>
294  %conv = extractelement <2 x i64> %0, i32 0
295  %add.ptr = getelementptr inbounds i64, ptr %__ptr, i64 %__offset
296  store i64 %conv, ptr %add.ptr, align 8
297  ret void
298}
299
300define void @vec_xst_trunc_ull(<1 x i128> %__vec, i64 %__offset, ptr nocapture %__ptr)  {
301; CHECK-LE-LABEL: vec_xst_trunc_ull:
302; CHECK-LE:       # %bb.0: # %entry
303; CHECK-LE-NEXT:    sldi r3, r5, 3
304; CHECK-LE-NEXT:    stxvrdx v2, r6, r3
305; CHECK-LE-NEXT:    blr
306;
307; CHECK-LINUXBE-LABEL: vec_xst_trunc_ull:
308; CHECK-LINUXBE:       # %bb.0: # %entry
309; CHECK-LINUXBE-NEXT:    sldi r3, r5, 3
310; CHECK-LINUXBE-NEXT:    stxsdx v2, r6, r3
311; CHECK-LINUXBE-NEXT:    blr
312;
313; CHECK-O0-LABEL: vec_xst_trunc_ull:
314; CHECK-O0:       # %bb.0: # %entry
315; CHECK-O0-NEXT:    mfvsrld r3, v2
316; CHECK-O0-NEXT:    sldi r4, r5, 3
317; CHECK-O0-NEXT:    add r4, r6, r4
318; CHECK-O0-NEXT:    std r3, 0(r4)
319; CHECK-O0-NEXT:    blr
320;
321; CHECK-AIXBE-LABEL: vec_xst_trunc_ull:
322; CHECK-AIXBE:       # %bb.0: # %entry
323; CHECK-AIXBE-NEXT:    sldi r3, r3, 3
324; CHECK-AIXBE-NEXT:    stxsdx v2, r4, r3
325; CHECK-AIXBE-NEXT:    blr
326entry:
327  %0 = bitcast <1 x i128> %__vec to <2 x i64>
328  %conv = extractelement <2 x i64> %0, i32 0
329  %add.ptr = getelementptr inbounds i64, ptr %__ptr, i64 %__offset
330  store i64 %conv, ptr %add.ptr, align 8
331  ret void
332}
333
334define dso_local <1 x i128> @vec_xl_zext(i64 %__offset, ptr nocapture readonly %__pointer) {
335; CHECK-LABEL: vec_xl_zext:
336; CHECK:       # %bb.0: # %entry
337; CHECK-NEXT:    lxvrbx v2, r4, r3
338; CHECK-NEXT:    blr
339entry:
340  %add.ptr = getelementptr inbounds i8, ptr %__pointer, i64 %__offset
341  %0 = load i8, ptr %add.ptr, align 1
342  %conv = zext i8 %0 to i128
343  %splat.splatinsert = insertelement <1 x i128> undef, i128 %conv, i32 0
344  ret <1 x i128> %splat.splatinsert
345}
346
347define dso_local <1 x i128> @vec_xl_zext_short(i64 %__offset, ptr nocapture readonly %__pointer) {
348; CHECK-LABEL: vec_xl_zext_short:
349; CHECK:       # %bb.0: # %entry
350; CHECK-NEXT:    sldi r3, r3, 1
351; CHECK-NEXT:    lxvrhx v2, r4, r3
352; CHECK-NEXT:    blr
353entry:
354  %add.ptr = getelementptr inbounds i16, ptr %__pointer, i64 %__offset
355  %0 = load i16, ptr %add.ptr, align 2
356  %conv = zext i16 %0 to i128
357  %splat.splatinsert = insertelement <1 x i128> undef, i128 %conv, i32 0
358  ret <1 x i128> %splat.splatinsert
359}
360
361define dso_local <1 x i128> @vec_xl_zext_word(i64 %__offset, ptr nocapture readonly %__pointer) {
362; CHECK-LABEL: vec_xl_zext_word:
363; CHECK:       # %bb.0: # %entry
364; CHECK-NEXT:    sldi r3, r3, 2
365; CHECK-NEXT:    lxvrwx v2, r4, r3
366; CHECK-NEXT:    blr
367entry:
368  %add.ptr = getelementptr inbounds i32, ptr %__pointer, i64 %__offset
369  %0 = load i32, ptr %add.ptr, align 4
370  %conv = zext i32 %0 to i128
371  %splat.splatinsert = insertelement <1 x i128> undef, i128 %conv, i32 0
372  ret <1 x i128> %splat.splatinsert
373}
374
375define dso_local <1 x i128> @vec_xl_zext_dw(i64 %__offset, ptr nocapture readonly %__pointer) {
376; CHECK-LABEL: vec_xl_zext_dw:
377; CHECK:       # %bb.0: # %entry
378; CHECK-NEXT:    sldi r3, r3, 3
379; CHECK-NEXT:    lxvrdx v2, r4, r3
380; CHECK-NEXT:    blr
381entry:
382  %add.ptr = getelementptr inbounds i64, ptr %__pointer, i64 %__offset
383  %0 = load i64, ptr %add.ptr, align 8
384  %conv = zext i64 %0 to i128
385  %splat.splatinsert = insertelement <1 x i128> undef, i128 %conv, i32 0
386  ret <1 x i128> %splat.splatinsert
387}
388
389define dso_local <1 x i128> @vec_xl_sext_b(i64 %offset, ptr %p) {
390; CHECK-LE-LABEL: vec_xl_sext_b:
391; CHECK-LE:       # %bb.0: # %entry
392; CHECK-LE-NEXT:    lbzx r3, r4, r3
393; CHECK-LE-NEXT:    extsb r3, r3
394; CHECK-LE-NEXT:    sradi r4, r3, 63
395; CHECK-LE-NEXT:    mtvsrdd v2, r4, r3
396; CHECK-LE-NEXT:    blr
397;
398; CHECK-LINUXBE-LABEL: vec_xl_sext_b:
399; CHECK-LINUXBE:       # %bb.0: # %entry
400; CHECK-LINUXBE-NEXT:    lbzx r3, r4, r3
401; CHECK-LINUXBE-NEXT:    extsb r3, r3
402; CHECK-LINUXBE-NEXT:    sradi r4, r3, 63
403; CHECK-LINUXBE-NEXT:    mtvsrdd v2, r4, r3
404; CHECK-LINUXBE-NEXT:    blr
405;
406; CHECK-O0-LABEL: vec_xl_sext_b:
407; CHECK-O0:       # %bb.0: # %entry
408; CHECK-O0-NEXT:    lbzx r3, r4, r3
409; CHECK-O0-NEXT:    extsb r4, r3
410; CHECK-O0-NEXT:    sradi r3, r4, 63
411; CHECK-O0-NEXT:    mtvsrdd v2, r3, r4
412; CHECK-O0-NEXT:    blr
413;
414; CHECK-AIXBE-LABEL: vec_xl_sext_b:
415; CHECK-AIXBE:       # %bb.0: # %entry
416; CHECK-AIXBE-NEXT:    lbzx r3, r4, r3
417; CHECK-AIXBE-NEXT:    extsb r3, r3
418; CHECK-AIXBE-NEXT:    sradi r4, r3, 63
419; CHECK-AIXBE-NEXT:    mtvsrdd v2, r4, r3
420; CHECK-AIXBE-NEXT:    blr
421entry:
422  %add.ptr = getelementptr inbounds i8, ptr %p, i64 %offset
423  %0 = load i8, ptr %add.ptr, align 1
424  %conv = sext i8 %0 to i128
425  %splat.splatinsert = insertelement <1 x i128> undef, i128 %conv, i32 0
426  ret <1 x i128> %splat.splatinsert
427}
428
429define dso_local <1 x i128> @vec_xl_sext_h(i64 %offset, ptr %p) {
430; CHECK-LE-LABEL: vec_xl_sext_h:
431; CHECK-LE:       # %bb.0: # %entry
432; CHECK-LE-NEXT:    sldi r3, r3, 1
433; CHECK-LE-NEXT:    lhax r3, r4, r3
434; CHECK-LE-NEXT:    sradi r4, r3, 63
435; CHECK-LE-NEXT:    mtvsrdd v2, r4, r3
436; CHECK-LE-NEXT:    blr
437;
438; CHECK-LINUXBE-LABEL: vec_xl_sext_h:
439; CHECK-LINUXBE:       # %bb.0: # %entry
440; CHECK-LINUXBE-NEXT:    sldi r3, r3, 1
441; CHECK-LINUXBE-NEXT:    lhax r3, r4, r3
442; CHECK-LINUXBE-NEXT:    sradi r4, r3, 63
443; CHECK-LINUXBE-NEXT:    mtvsrdd v2, r4, r3
444; CHECK-LINUXBE-NEXT:    blr
445;
446; CHECK-O0-LABEL: vec_xl_sext_h:
447; CHECK-O0:       # %bb.0: # %entry
448; CHECK-O0-NEXT:    sldi r3, r3, 1
449; CHECK-O0-NEXT:    lhax r4, r4, r3
450; CHECK-O0-NEXT:    sradi r3, r4, 63
451; CHECK-O0-NEXT:    mtvsrdd v2, r3, r4
452; CHECK-O0-NEXT:    blr
453;
454; CHECK-AIXBE-LABEL: vec_xl_sext_h:
455; CHECK-AIXBE:       # %bb.0: # %entry
456; CHECK-AIXBE-NEXT:    sldi r3, r3, 1
457; CHECK-AIXBE-NEXT:    lhax r3, r4, r3
458; CHECK-AIXBE-NEXT:    sradi r4, r3, 63
459; CHECK-AIXBE-NEXT:    mtvsrdd v2, r4, r3
460; CHECK-AIXBE-NEXT:    blr
461entry:
462  %add.ptr = getelementptr inbounds i16, ptr %p, i64 %offset
463  %0 = load i16, ptr %add.ptr, align 2
464  %conv = sext i16 %0 to i128
465  %splat.splatinsert = insertelement <1 x i128> undef, i128 %conv, i32 0
466  ret <1 x i128> %splat.splatinsert
467}
468
469define dso_local <1 x i128> @vec_xl_sext_w(i64 %offset, ptr %p) {
470; CHECK-LE-LABEL: vec_xl_sext_w:
471; CHECK-LE:       # %bb.0: # %entry
472; CHECK-LE-NEXT:    sldi r3, r3, 2
473; CHECK-LE-NEXT:    lwax r3, r4, r3
474; CHECK-LE-NEXT:    sradi r4, r3, 63
475; CHECK-LE-NEXT:    mtvsrdd v2, r4, r3
476; CHECK-LE-NEXT:    blr
477;
478; CHECK-LINUXBE-LABEL: vec_xl_sext_w:
479; CHECK-LINUXBE:       # %bb.0: # %entry
480; CHECK-LINUXBE-NEXT:    sldi r3, r3, 2
481; CHECK-LINUXBE-NEXT:    lwax r3, r4, r3
482; CHECK-LINUXBE-NEXT:    sradi r4, r3, 63
483; CHECK-LINUXBE-NEXT:    mtvsrdd v2, r4, r3
484; CHECK-LINUXBE-NEXT:    blr
485;
486; CHECK-O0-LABEL: vec_xl_sext_w:
487; CHECK-O0:       # %bb.0: # %entry
488; CHECK-O0-NEXT:    sldi r3, r3, 2
489; CHECK-O0-NEXT:    lwax r4, r4, r3
490; CHECK-O0-NEXT:    sradi r3, r4, 63
491; CHECK-O0-NEXT:    mtvsrdd v2, r3, r4
492; CHECK-O0-NEXT:    blr
493;
494; CHECK-AIXBE-LABEL: vec_xl_sext_w:
495; CHECK-AIXBE:       # %bb.0: # %entry
496; CHECK-AIXBE-NEXT:    sldi r3, r3, 2
497; CHECK-AIXBE-NEXT:    lwax r3, r4, r3
498; CHECK-AIXBE-NEXT:    sradi r4, r3, 63
499; CHECK-AIXBE-NEXT:    mtvsrdd v2, r4, r3
500; CHECK-AIXBE-NEXT:    blr
501entry:
502  %add.ptr = getelementptr inbounds i32, ptr %p, i64 %offset
503  %0 = load i32, ptr %add.ptr, align 4
504  %conv = sext i32 %0 to i128
505  %splat.splatinsert = insertelement <1 x i128> undef, i128 %conv, i32 0
506  ret <1 x i128> %splat.splatinsert
507}
508
509define dso_local <1 x i128> @vec_xl_sext_d(i64 %offset, ptr %p) {
510; CHECK-LE-LABEL: vec_xl_sext_d:
511; CHECK-LE:       # %bb.0: # %entry
512; CHECK-LE-NEXT:    sldi r3, r3, 3
513; CHECK-LE-NEXT:    ldx r3, r4, r3
514; CHECK-LE-NEXT:    sradi r4, r3, 63
515; CHECK-LE-NEXT:    mtvsrdd v2, r4, r3
516; CHECK-LE-NEXT:    blr
517;
518; CHECK-LINUXBE-LABEL: vec_xl_sext_d:
519; CHECK-LINUXBE:       # %bb.0: # %entry
520; CHECK-LINUXBE-NEXT:    sldi r3, r3, 3
521; CHECK-LINUXBE-NEXT:    ldx r3, r4, r3
522; CHECK-LINUXBE-NEXT:    sradi r4, r3, 63
523; CHECK-LINUXBE-NEXT:    mtvsrdd v2, r4, r3
524; CHECK-LINUXBE-NEXT:    blr
525;
526; CHECK-O0-LABEL: vec_xl_sext_d:
527; CHECK-O0:       # %bb.0: # %entry
528; CHECK-O0-NEXT:    sldi r3, r3, 3
529; CHECK-O0-NEXT:    ldx r4, r4, r3
530; CHECK-O0-NEXT:    sradi r3, r4, 63
531; CHECK-O0-NEXT:    mtvsrdd v2, r3, r4
532; CHECK-O0-NEXT:    blr
533;
534; CHECK-AIXBE-LABEL: vec_xl_sext_d:
535; CHECK-AIXBE:       # %bb.0: # %entry
536; CHECK-AIXBE-NEXT:    sldi r3, r3, 3
537; CHECK-AIXBE-NEXT:    ldx r3, r4, r3
538; CHECK-AIXBE-NEXT:    sradi r4, r3, 63
539; CHECK-AIXBE-NEXT:    mtvsrdd v2, r4, r3
540; CHECK-AIXBE-NEXT:    blr
541entry:
542  %add.ptr = getelementptr inbounds i64, ptr %p, i64 %offset
543  %0 = load i64, ptr %add.ptr, align 8
544  %conv = sext i64 %0 to i128
545  %splat.splatinsert = insertelement <1 x i128> undef, i128 %conv, i32 0
546  ret <1 x i128> %splat.splatinsert
547}
548