Gitiles
Code Review Sign In
gerrit-public.fairphone.software / toolchain / llvm-project / bdbf839a3bc52cfafcf799f1d06c9f3841f2b0a2 / . / llvm / test / CodeGen / X86 / vec_cmp_uint-128.ll
blob: 305fd80b7c40ff14d715d2c804df97525dfb0039 [file] [log] [blame]
Simon Pilgrimbdbf8392015-10-10 22:21:05 +0000[diff] [blame^]1; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=ALL --check-prefix=SSE --check-prefix=SSE2
2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=ALL --check-prefix=SSE --check-prefix=SSE41
3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.2 | FileCheck %s --check-prefix=ALL --check-prefix=SSE --check-prefix=SSE42
4; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX1
5; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX2
6; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx,+xop | FileCheck %s --check-prefix=ALL --check-prefix=XOP --check-prefix=XOPAVX1
7; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2,+xop | FileCheck %s --check-prefix=ALL --check-prefix=XOP --check-prefix=XOPAVX2
8; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX512 --check-prefix=AVX512F
9; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX512 --check-prefix=AVX512BW
10
11;
12; Equal
13;
14
15define <2 x i64> @eq_v2i64(<2 x i64> %a, <2 x i64> %b) nounwind {
16; SSE2-LABEL: eq_v2i64:
17; SSE2: # BB#0:
18; SSE2-NEXT: pcmpeqd %xmm1, %xmm0
19; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,0,3,2]
20; SSE2-NEXT: pand %xmm1, %xmm0
21; SSE2-NEXT: retq
22;
23; SSE41-LABEL: eq_v2i64:
24; SSE41: # BB#0:
25; SSE41-NEXT: pcmpeqq %xmm1, %xmm0
26; SSE41-NEXT: retq
27;
28; SSE42-LABEL: eq_v2i64:
29; SSE42: # BB#0:
30; SSE42-NEXT: pcmpeqq %xmm1, %xmm0
31; SSE42-NEXT: retq
32;
33; AVX-LABEL: eq_v2i64:
34; AVX: # BB#0:
35; AVX-NEXT: vpcmpeqq %xmm1, %xmm0, %xmm0
36; AVX-NEXT: retq
37;
38; XOP-LABEL: eq_v2i64:
39; XOP: # BB#0:
40; XOP-NEXT: vpcmpeqq %xmm1, %xmm0, %xmm0
41; XOP-NEXT: retq
42 %1 = icmp eq <2 x i64> %a, %b
43 %2 = sext <2 x i1> %1 to <2 x i64>
44 ret <2 x i64> %2
45}
46
47define <4 x i32> @eq_v4i32(<4 x i32> %a, <4 x i32> %b) nounwind {
48; SSE-LABEL: eq_v4i32:
49; SSE: # BB#0:
50; SSE-NEXT: pcmpeqd %xmm1, %xmm0
51; SSE-NEXT: retq
52;
53; AVX-LABEL: eq_v4i32:
54; AVX: # BB#0:
55; AVX-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm0
56; AVX-NEXT: retq
57;
58; XOP-LABEL: eq_v4i32:
59; XOP: # BB#0:
60; XOP-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm0
61; XOP-NEXT: retq
62 %1 = icmp eq <4 x i32> %a, %b
63 %2 = sext <4 x i1> %1 to <4 x i32>
64 ret <4 x i32> %2
65}
66
67define <8 x i16> @eq_v8i16(<8 x i16> %a, <8 x i16> %b) nounwind {
68; SSE-LABEL: eq_v8i16:
69; SSE: # BB#0:
70; SSE-NEXT: pcmpeqw %xmm1, %xmm0
71; SSE-NEXT: retq
72;
73; AVX-LABEL: eq_v8i16:
74; AVX: # BB#0:
75; AVX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
76; AVX-NEXT: retq
77;
78; XOP-LABEL: eq_v8i16:
79; XOP: # BB#0:
80; XOP-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
81; XOP-NEXT: retq
82 %1 = icmp eq <8 x i16> %a, %b
83 %2 = sext <8 x i1> %1 to <8 x i16>
84 ret <8 x i16> %2
85}
86
87define <16 x i8> @eq_v16i8(<16 x i8> %a, <16 x i8> %b) nounwind {
88; SSE-LABEL: eq_v16i8:
89; SSE: # BB#0:
90; SSE-NEXT: pcmpeqb %xmm1, %xmm0
91; SSE-NEXT: retq
92;
93; AVX-LABEL: eq_v16i8:
94; AVX: # BB#0:
95; AVX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
96; AVX-NEXT: retq
97;
98; XOP-LABEL: eq_v16i8:
99; XOP: # BB#0:
100; XOP-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
101; XOP-NEXT: retq
102 %1 = icmp eq <16 x i8> %a, %b
103 %2 = sext <16 x i1> %1 to <16 x i8>
104 ret <16 x i8> %2
105}
106
107;
108; Not Equal
109;
110
111define <2 x i64> @ne_v2i64(<2 x i64> %a, <2 x i64> %b) nounwind {
112; SSE2-LABEL: ne_v2i64:
113; SSE2: # BB#0:
114; SSE2-NEXT: pcmpeqd %xmm1, %xmm0
115; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,0,3,2]
116; SSE2-NEXT: pand %xmm1, %xmm0
117; SSE2-NEXT: pcmpeqd %xmm1, %xmm1
118; SSE2-NEXT: pxor %xmm1, %xmm0
119; SSE2-NEXT: retq
120;
121; SSE41-LABEL: ne_v2i64:
122; SSE41: # BB#0:
123; SSE41-NEXT: pcmpeqq %xmm1, %xmm0
124; SSE41-NEXT: pcmpeqd %xmm1, %xmm1
125; SSE41-NEXT: pxor %xmm1, %xmm0
126; SSE41-NEXT: retq
127;
128; SSE42-LABEL: ne_v2i64:
129; SSE42: # BB#0:
130; SSE42-NEXT: pcmpeqq %xmm1, %xmm0
131; SSE42-NEXT: pcmpeqd %xmm1, %xmm1
132; SSE42-NEXT: pxor %xmm1, %xmm0
133; SSE42-NEXT: retq
134;
135; AVX-LABEL: ne_v2i64:
136; AVX: # BB#0:
137; AVX-NEXT: vpcmpeqq %xmm1, %xmm0, %xmm0
138; AVX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
139; AVX-NEXT: vpxor %xmm1, %xmm0, %xmm0
140; AVX-NEXT: retq
141;
142; XOP-LABEL: ne_v2i64:
143; XOP: # BB#0:
144; XOP-NEXT: vpcmpeqq %xmm1, %xmm0, %xmm0
145; XOP-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
146; XOP-NEXT: vpxor %xmm1, %xmm0, %xmm0
147; XOP-NEXT: retq
148 %1 = icmp ne <2 x i64> %a, %b
149 %2 = sext <2 x i1> %1 to <2 x i64>
150 ret <2 x i64> %2
151}
152
153define <4 x i32> @ne_v4i32(<4 x i32> %a, <4 x i32> %b) nounwind {
154; SSE-LABEL: ne_v4i32:
155; SSE: # BB#0:
156; SSE-NEXT: pcmpeqd %xmm1, %xmm0
157; SSE-NEXT: pcmpeqd %xmm1, %xmm1
158; SSE-NEXT: pxor %xmm1, %xmm0
159; SSE-NEXT: retq
160;
161; AVX-LABEL: ne_v4i32:
162; AVX: # BB#0:
163; AVX-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm0
164; AVX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
165; AVX-NEXT: vpxor %xmm1, %xmm0, %xmm0
166; AVX-NEXT: retq
167;
168; XOP-LABEL: ne_v4i32:
169; XOP: # BB#0:
170; XOP-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm0
171; XOP-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
172; XOP-NEXT: vpxor %xmm1, %xmm0, %xmm0
173; XOP-NEXT: retq
174 %1 = icmp ne <4 x i32> %a, %b
175 %2 = sext <4 x i1> %1 to <4 x i32>
176 ret <4 x i32> %2
177}
178
179define <8 x i16> @ne_v8i16(<8 x i16> %a, <8 x i16> %b) nounwind {
180; SSE-LABEL: ne_v8i16:
181; SSE: # BB#0:
182; SSE-NEXT: pcmpeqw %xmm1, %xmm0
183; SSE-NEXT: pcmpeqd %xmm1, %xmm1
184; SSE-NEXT: pxor %xmm1, %xmm0
185; SSE-NEXT: retq
186;
187; AVX-LABEL: ne_v8i16:
188; AVX: # BB#0:
189; AVX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
190; AVX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
191; AVX-NEXT: vpxor %xmm1, %xmm0, %xmm0
192; AVX-NEXT: retq
193;
194; XOP-LABEL: ne_v8i16:
195; XOP: # BB#0:
196; XOP-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
197; XOP-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
198; XOP-NEXT: vpxor %xmm1, %xmm0, %xmm0
199; XOP-NEXT: retq
200 %1 = icmp ne <8 x i16> %a, %b
201 %2 = sext <8 x i1> %1 to <8 x i16>
202 ret <8 x i16> %2
203}
204
205define <16 x i8> @ne_v16i8(<16 x i8> %a, <16 x i8> %b) nounwind {
206; SSE-LABEL: ne_v16i8:
207; SSE: # BB#0:
208; SSE-NEXT: pcmpeqb %xmm1, %xmm0
209; SSE-NEXT: pcmpeqd %xmm1, %xmm1
210; SSE-NEXT: pxor %xmm1, %xmm0
211; SSE-NEXT: retq
212;
213; AVX-LABEL: ne_v16i8:
214; AVX: # BB#0:
215; AVX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
216; AVX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
217; AVX-NEXT: vpxor %xmm1, %xmm0, %xmm0
218; AVX-NEXT: retq
219;
220; XOP-LABEL: ne_v16i8:
221; XOP: # BB#0:
222; XOP-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
223; XOP-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
224; XOP-NEXT: vpxor %xmm1, %xmm0, %xmm0
225; XOP-NEXT: retq
226 %1 = icmp ne <16 x i8> %a, %b
227 %2 = sext <16 x i1> %1 to <16 x i8>
228 ret <16 x i8> %2
229}
230
231;
232; Greater Than Or Equal
233;
234
235define <2 x i64> @ge_v2i64(<2 x i64> %a, <2 x i64> %b) nounwind {
236; SSE2-LABEL: ge_v2i64:
237; SSE2: # BB#0:
238; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
239; SSE2-NEXT: pxor %xmm2, %xmm0
240; SSE2-NEXT: pxor %xmm2, %xmm1
241; SSE2-NEXT: movdqa %xmm1, %xmm2
242; SSE2-NEXT: pcmpgtd %xmm0, %xmm2
243; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
244; SSE2-NEXT: pcmpeqd %xmm0, %xmm1
245; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
246; SSE2-NEXT: pand %xmm3, %xmm0
247; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
248; SSE2-NEXT: por %xmm0, %xmm1
249; SSE2-NEXT: pcmpeqd %xmm0, %xmm0
250; SSE2-NEXT: pxor %xmm1, %xmm0
251; SSE2-NEXT: retq
252;
253; SSE41-LABEL: ge_v2i64:
254; SSE41: # BB#0:
255; SSE41-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
256; SSE41-NEXT: pxor %xmm2, %xmm0
257; SSE41-NEXT: pxor %xmm2, %xmm1
258; SSE41-NEXT: movdqa %xmm1, %xmm2
259; SSE41-NEXT: pcmpgtd %xmm0, %xmm2
260; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
261; SSE41-NEXT: pcmpeqd %xmm0, %xmm1
262; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
263; SSE41-NEXT: pand %xmm3, %xmm0
264; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
265; SSE41-NEXT: por %xmm0, %xmm1
266; SSE41-NEXT: pcmpeqd %xmm0, %xmm0
267; SSE41-NEXT: pxor %xmm1, %xmm0
268; SSE41-NEXT: retq
269;
270; SSE42-LABEL: ge_v2i64:
271; SSE42: # BB#0:
272; SSE42-NEXT: movdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
273; SSE42-NEXT: pxor %xmm2, %xmm0
274; SSE42-NEXT: pxor %xmm1, %xmm2
275; SSE42-NEXT: pcmpgtq %xmm0, %xmm2
276; SSE42-NEXT: pcmpeqd %xmm0, %xmm0
277; SSE42-NEXT: pxor %xmm2, %xmm0
278; SSE42-NEXT: retq
279;
280; AVX-LABEL: ge_v2i64:
281; AVX: # BB#0:
282; AVX-NEXT: vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
283; AVX-NEXT: vpxor %xmm2, %xmm0, %xmm0
284; AVX-NEXT: vpxor %xmm2, %xmm1, %xmm1
285; AVX-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0
286; AVX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
287; AVX-NEXT: vpxor %xmm1, %xmm0, %xmm0
288; AVX-NEXT: retq
289;
290; XOP-LABEL: ge_v2i64:
291; XOP: # BB#0:
292; XOP-NEXT: vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
293; XOP-NEXT: vpxor %xmm2, %xmm0, %xmm0
294; XOP-NEXT: vpxor %xmm2, %xmm1, %xmm1
295; XOP-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0
296; XOP-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
297; XOP-NEXT: vpxor %xmm1, %xmm0, %xmm0
298; XOP-NEXT: retq
299 %1 = icmp uge <2 x i64> %a, %b
300 %2 = sext <2 x i1> %1 to <2 x i64>
301 ret <2 x i64> %2
302}
303
304define <4 x i32> @ge_v4i32(<4 x i32> %a, <4 x i32> %b) nounwind {
305; SSE2-LABEL: ge_v4i32:
306; SSE2: # BB#0:
307; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
308; SSE2-NEXT: pxor %xmm2, %xmm0
309; SSE2-NEXT: pxor %xmm1, %xmm2
310; SSE2-NEXT: pcmpgtd %xmm0, %xmm2
311; SSE2-NEXT: pcmpeqd %xmm0, %xmm0
312; SSE2-NEXT: pxor %xmm2, %xmm0
313; SSE2-NEXT: retq
314;
315; SSE41-LABEL: ge_v4i32:
316; SSE41: # BB#0:
317; SSE41-NEXT: pmaxud %xmm0, %xmm1
318; SSE41-NEXT: pcmpeqd %xmm1, %xmm0
319; SSE41-NEXT: retq
320;
321; SSE42-LABEL: ge_v4i32:
322; SSE42: # BB#0:
323; SSE42-NEXT: pmaxud %xmm0, %xmm1
324; SSE42-NEXT: pcmpeqd %xmm1, %xmm0
325; SSE42-NEXT: retq
326;
327; AVX-LABEL: ge_v4i32:
328; AVX: # BB#0:
329; AVX-NEXT: vpmaxud %xmm1, %xmm0, %xmm1
330; AVX-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm0
331; AVX-NEXT: retq
332;
333; XOP-LABEL: ge_v4i32:
334; XOP: # BB#0:
335; XOP-NEXT: vpmaxud %xmm1, %xmm0, %xmm1
336; XOP-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm0
337; XOP-NEXT: retq
338 %1 = icmp uge <4 x i32> %a, %b
339 %2 = sext <4 x i1> %1 to <4 x i32>
340 ret <4 x i32> %2
341}
342
343define <8 x i16> @ge_v8i16(<8 x i16> %a, <8 x i16> %b) nounwind {
344; SSE2-LABEL: ge_v8i16:
345; SSE2: # BB#0:
346; SSE2-NEXT: psubusw %xmm0, %xmm1
347; SSE2-NEXT: pxor %xmm0, %xmm0
348; SSE2-NEXT: pcmpeqw %xmm1, %xmm0
349; SSE2-NEXT: retq
350;
351; SSE41-LABEL: ge_v8i16:
352; SSE41: # BB#0:
353; SSE41-NEXT: pmaxuw %xmm0, %xmm1
354; SSE41-NEXT: pcmpeqw %xmm1, %xmm0
355; SSE41-NEXT: retq
356;
357; SSE42-LABEL: ge_v8i16:
358; SSE42: # BB#0:
359; SSE42-NEXT: pmaxuw %xmm0, %xmm1
360; SSE42-NEXT: pcmpeqw %xmm1, %xmm0
361; SSE42-NEXT: retq
362;
363; AVX-LABEL: ge_v8i16:
364; AVX: # BB#0:
365; AVX-NEXT: vpmaxuw %xmm1, %xmm0, %xmm1
366; AVX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
367; AVX-NEXT: retq
368;
369; XOP-LABEL: ge_v8i16:
370; XOP: # BB#0:
371; XOP-NEXT: vpmaxuw %xmm1, %xmm0, %xmm1
372; XOP-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
373; XOP-NEXT: retq
374 %1 = icmp uge <8 x i16> %a, %b
375 %2 = sext <8 x i1> %1 to <8 x i16>
376 ret <8 x i16> %2
377}
378
379define <16 x i8> @ge_v16i8(<16 x i8> %a, <16 x i8> %b) nounwind {
380; SSE-LABEL: ge_v16i8:
381; SSE: # BB#0:
382; SSE-NEXT: pmaxub %xmm0, %xmm1
383; SSE-NEXT: pcmpeqb %xmm1, %xmm0
384; SSE-NEXT: retq
385;
386; AVX-LABEL: ge_v16i8:
387; AVX: # BB#0:
388; AVX-NEXT: vpmaxub %xmm1, %xmm0, %xmm1
389; AVX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
390; AVX-NEXT: retq
391;
392; XOP-LABEL: ge_v16i8:
393; XOP: # BB#0:
394; XOP-NEXT: vpmaxub %xmm1, %xmm0, %xmm1
395; XOP-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
396; XOP-NEXT: retq
397 %1 = icmp uge <16 x i8> %a, %b
398 %2 = sext <16 x i1> %1 to <16 x i8>
399 ret <16 x i8> %2
400}
401
402;
403; Greater Than
404;
405
406define <2 x i64> @gt_v2i64(<2 x i64> %a, <2 x i64> %b) nounwind {
407; SSE2-LABEL: gt_v2i64:
408; SSE2: # BB#0:
409; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
410; SSE2-NEXT: pxor %xmm2, %xmm1
411; SSE2-NEXT: pxor %xmm2, %xmm0
412; SSE2-NEXT: movdqa %xmm0, %xmm2
413; SSE2-NEXT: pcmpgtd %xmm1, %xmm2
414; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
415; SSE2-NEXT: pcmpeqd %xmm1, %xmm0
416; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
417; SSE2-NEXT: pand %xmm3, %xmm1
418; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
419; SSE2-NEXT: por %xmm1, %xmm0
420; SSE2-NEXT: retq
421;
422; SSE41-LABEL: gt_v2i64:
423; SSE41: # BB#0:
424; SSE41-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
425; SSE41-NEXT: pxor %xmm2, %xmm1
426; SSE41-NEXT: pxor %xmm2, %xmm0
427; SSE41-NEXT: movdqa %xmm0, %xmm2
428; SSE41-NEXT: pcmpgtd %xmm1, %xmm2
429; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
430; SSE41-NEXT: pcmpeqd %xmm1, %xmm0
431; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
432; SSE41-NEXT: pand %xmm3, %xmm1
433; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
434; SSE41-NEXT: por %xmm1, %xmm0
435; SSE41-NEXT: retq
436;
437; SSE42-LABEL: gt_v2i64:
438; SSE42: # BB#0:
439; SSE42-NEXT: movdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
440; SSE42-NEXT: pxor %xmm2, %xmm1
441; SSE42-NEXT: pxor %xmm2, %xmm0
442; SSE42-NEXT: pcmpgtq %xmm1, %xmm0
443; SSE42-NEXT: retq
444;
445; AVX-LABEL: gt_v2i64:
446; AVX: # BB#0:
447; AVX-NEXT: vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
448; AVX-NEXT: vpxor %xmm2, %xmm1, %xmm1
449; AVX-NEXT: vpxor %xmm2, %xmm0, %xmm0
450; AVX-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm0
451; AVX-NEXT: retq
452;
453; XOP-LABEL: gt_v2i64:
454; XOP: # BB#0:
455; XOP-NEXT: vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
456; XOP-NEXT: vpxor %xmm2, %xmm1, %xmm1
457; XOP-NEXT: vpxor %xmm2, %xmm0, %xmm0
458; XOP-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm0
459; XOP-NEXT: retq
460 %1 = icmp ugt <2 x i64> %a, %b
461 %2 = sext <2 x i1> %1 to <2 x i64>
462 ret <2 x i64> %2
463}
464
465define <4 x i32> @gt_v4i32(<4 x i32> %a, <4 x i32> %b) nounwind {
466; SSE-LABEL: gt_v4i32:
467; SSE: # BB#0:
468; SSE-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
469; SSE-NEXT: pxor %xmm2, %xmm1
470; SSE-NEXT: pxor %xmm2, %xmm0
471; SSE-NEXT: pcmpgtd %xmm1, %xmm0
472; SSE-NEXT: retq
473;
474; AVX1-LABEL: gt_v4i32:
475; AVX1: # BB#0:
476; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
477; AVX1-NEXT: vpxor %xmm2, %xmm1, %xmm1
478; AVX1-NEXT: vpxor %xmm2, %xmm0, %xmm0
479; AVX1-NEXT: vpcmpgtd %xmm1, %xmm0, %xmm0
480; AVX1-NEXT: retq
481;
482; AVX2-LABEL: gt_v4i32:
483; AVX2: # BB#0:
484; AVX2-NEXT: vpbroadcastd {{.*}}(%rip), %xmm2
485; AVX2-NEXT: vpxor %xmm2, %xmm1, %xmm1
486; AVX2-NEXT: vpxor %xmm2, %xmm0, %xmm0
487; AVX2-NEXT: vpcmpgtd %xmm1, %xmm0, %xmm0
488; AVX2-NEXT: retq
489;
490; XOPAVX1-LABEL: gt_v4i32:
491; XOPAVX1: # BB#0:
492; XOPAVX1-NEXT: vmovdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
493; XOPAVX1-NEXT: vpxor %xmm2, %xmm1, %xmm1
494; XOPAVX1-NEXT: vpxor %xmm2, %xmm0, %xmm0
495; XOPAVX1-NEXT: vpcmpgtd %xmm1, %xmm0, %xmm0
496; XOPAVX1-NEXT: retq
497;
498; XOPAVX2-LABEL: gt_v4i32:
499; XOPAVX2: # BB#0:
500; XOPAVX2-NEXT: vpbroadcastd {{.*}}(%rip), %xmm2
501; XOPAVX2-NEXT: vpxor %xmm2, %xmm1, %xmm1
502; XOPAVX2-NEXT: vpxor %xmm2, %xmm0, %xmm0
503; XOPAVX2-NEXT: vpcmpgtd %xmm1, %xmm0, %xmm0
504; XOPAVX2-NEXT: retq
505;
506; AVX512-LABEL: gt_v4i32:
507; AVX512: # BB#0:
508; AVX512-NEXT: vpbroadcastd {{.*}}(%rip), %xmm2
509; AVX512-NEXT: vpxor %xmm2, %xmm1, %xmm1
510; AVX512-NEXT: vpxor %xmm2, %xmm0, %xmm0
511; AVX512-NEXT: vpcmpgtd %xmm1, %xmm0, %xmm0
512; AVX512-NEXT: retq
513 %1 = icmp ugt <4 x i32> %a, %b
514 %2 = sext <4 x i1> %1 to <4 x i32>
515 ret <4 x i32> %2
516}
517
518define <8 x i16> @gt_v8i16(<8 x i16> %a, <8 x i16> %b) nounwind {
519; SSE-LABEL: gt_v8i16:
520; SSE: # BB#0:
521; SSE-NEXT: movdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
522; SSE-NEXT: pxor %xmm2, %xmm1
523; SSE-NEXT: pxor %xmm2, %xmm0
524; SSE-NEXT: pcmpgtw %xmm1, %xmm0
525; SSE-NEXT: retq
526;
527; AVX-LABEL: gt_v8i16:
528; AVX: # BB#0:
529; AVX-NEXT: vmovdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
530; AVX-NEXT: vpxor %xmm2, %xmm1, %xmm1
531; AVX-NEXT: vpxor %xmm2, %xmm0, %xmm0
532; AVX-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0
533; AVX-NEXT: retq
534;
535; XOP-LABEL: gt_v8i16:
536; XOP: # BB#0:
537; XOP-NEXT: vmovdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
538; XOP-NEXT: vpxor %xmm2, %xmm1, %xmm1
539; XOP-NEXT: vpxor %xmm2, %xmm0, %xmm0
540; XOP-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0
541; XOP-NEXT: retq
542 %1 = icmp ugt <8 x i16> %a, %b
543 %2 = sext <8 x i1> %1 to <8 x i16>
544 ret <8 x i16> %2
545}
546
547define <16 x i8> @gt_v16i8(<16 x i8> %a, <16 x i8> %b) nounwind {
548; SSE-LABEL: gt_v16i8:
549; SSE: # BB#0:
550; SSE-NEXT: movdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
551; SSE-NEXT: pxor %xmm2, %xmm1
552; SSE-NEXT: pxor %xmm2, %xmm0
553; SSE-NEXT: pcmpgtb %xmm1, %xmm0
554; SSE-NEXT: retq
555;
556; AVX-LABEL: gt_v16i8:
557; AVX: # BB#0:
558; AVX-NEXT: vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
559; AVX-NEXT: vpxor %xmm2, %xmm1, %xmm1
560; AVX-NEXT: vpxor %xmm2, %xmm0, %xmm0
561; AVX-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm0
562; AVX-NEXT: retq
563;
564; XOP-LABEL: gt_v16i8:
565; XOP: # BB#0:
566; XOP-NEXT: vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
567; XOP-NEXT: vpxor %xmm2, %xmm1, %xmm1
568; XOP-NEXT: vpxor %xmm2, %xmm0, %xmm0
569; XOP-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm0
570; XOP-NEXT: retq
571 %1 = icmp ugt <16 x i8> %a, %b
572 %2 = sext <16 x i1> %1 to <16 x i8>
573 ret <16 x i8> %2
574}
575
576;
577; Less Than Or Equal
578;
579
580define <2 x i64> @le_v2i64(<2 x i64> %a, <2 x i64> %b) nounwind {
581; SSE2-LABEL: le_v2i64:
582; SSE2: # BB#0:
583; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
584; SSE2-NEXT: pxor %xmm2, %xmm1
585; SSE2-NEXT: pxor %xmm2, %xmm0
586; SSE2-NEXT: movdqa %xmm0, %xmm2
587; SSE2-NEXT: pcmpgtd %xmm1, %xmm2
588; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
589; SSE2-NEXT: pcmpeqd %xmm1, %xmm0
590; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
591; SSE2-NEXT: pand %xmm3, %xmm0
592; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
593; SSE2-NEXT: por %xmm0, %xmm1
594; SSE2-NEXT: pcmpeqd %xmm0, %xmm0
595; SSE2-NEXT: pxor %xmm1, %xmm0
596; SSE2-NEXT: retq
597;
598; SSE41-LABEL: le_v2i64:
599; SSE41: # BB#0:
600; SSE41-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
601; SSE41-NEXT: pxor %xmm2, %xmm1
602; SSE41-NEXT: pxor %xmm2, %xmm0
603; SSE41-NEXT: movdqa %xmm0, %xmm2
604; SSE41-NEXT: pcmpgtd %xmm1, %xmm2
605; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
606; SSE41-NEXT: pcmpeqd %xmm1, %xmm0
607; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
608; SSE41-NEXT: pand %xmm3, %xmm0
609; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
610; SSE41-NEXT: por %xmm0, %xmm1
611; SSE41-NEXT: pcmpeqd %xmm0, %xmm0
612; SSE41-NEXT: pxor %xmm1, %xmm0
613; SSE41-NEXT: retq
614;
615; SSE42-LABEL: le_v2i64:
616; SSE42: # BB#0:
617; SSE42-NEXT: movdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
618; SSE42-NEXT: pxor %xmm2, %xmm1
619; SSE42-NEXT: pxor %xmm2, %xmm0
620; SSE42-NEXT: pcmpgtq %xmm1, %xmm0
621; SSE42-NEXT: pcmpeqd %xmm1, %xmm1
622; SSE42-NEXT: pxor %xmm1, %xmm0
623; SSE42-NEXT: retq
624;
625; AVX-LABEL: le_v2i64:
626; AVX: # BB#0:
627; AVX-NEXT: vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
628; AVX-NEXT: vpxor %xmm2, %xmm1, %xmm1
629; AVX-NEXT: vpxor %xmm2, %xmm0, %xmm0
630; AVX-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm0
631; AVX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
632; AVX-NEXT: vpxor %xmm1, %xmm0, %xmm0
633; AVX-NEXT: retq
634;
635; XOP-LABEL: le_v2i64:
636; XOP: # BB#0:
637; XOP-NEXT: vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
638; XOP-NEXT: vpxor %xmm2, %xmm1, %xmm1
639; XOP-NEXT: vpxor %xmm2, %xmm0, %xmm0
640; XOP-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm0
641; XOP-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
642; XOP-NEXT: vpxor %xmm1, %xmm0, %xmm0
643; XOP-NEXT: retq
644 %1 = icmp ule <2 x i64> %a, %b
645 %2 = sext <2 x i1> %1 to <2 x i64>
646 ret <2 x i64> %2
647}
648
649define <4 x i32> @le_v4i32(<4 x i32> %a, <4 x i32> %b) nounwind {
650; SSE2-LABEL: le_v4i32:
651; SSE2: # BB#0:
652; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
653; SSE2-NEXT: pxor %xmm2, %xmm1
654; SSE2-NEXT: pxor %xmm2, %xmm0
655; SSE2-NEXT: pcmpgtd %xmm1, %xmm0
656; SSE2-NEXT: pcmpeqd %xmm1, %xmm1
657; SSE2-NEXT: pxor %xmm1, %xmm0
658; SSE2-NEXT: retq
659;
660; SSE41-LABEL: le_v4i32:
661; SSE41: # BB#0:
662; SSE41-NEXT: pminud %xmm0, %xmm1
663; SSE41-NEXT: pcmpeqd %xmm1, %xmm0
664; SSE41-NEXT: retq
665;
666; SSE42-LABEL: le_v4i32:
667; SSE42: # BB#0:
668; SSE42-NEXT: pminud %xmm0, %xmm1
669; SSE42-NEXT: pcmpeqd %xmm1, %xmm0
670; SSE42-NEXT: retq
671;
672; AVX-LABEL: le_v4i32:
673; AVX: # BB#0:
674; AVX-NEXT: vpminud %xmm1, %xmm0, %xmm1
675; AVX-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm0
676; AVX-NEXT: retq
677;
678; XOP-LABEL: le_v4i32:
679; XOP: # BB#0:
680; XOP-NEXT: vpminud %xmm1, %xmm0, %xmm1
681; XOP-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm0
682; XOP-NEXT: retq
683 %1 = icmp ule <4 x i32> %a, %b
684 %2 = sext <4 x i1> %1 to <4 x i32>
685 ret <4 x i32> %2
686}
687
688define <8 x i16> @le_v8i16(<8 x i16> %a, <8 x i16> %b) nounwind {
689; SSE2-LABEL: le_v8i16:
690; SSE2: # BB#0:
691; SSE2-NEXT: psubusw %xmm1, %xmm0
692; SSE2-NEXT: pxor %xmm1, %xmm1
693; SSE2-NEXT: pcmpeqw %xmm1, %xmm0
694; SSE2-NEXT: retq
695;
696; SSE41-LABEL: le_v8i16:
697; SSE41: # BB#0:
698; SSE41-NEXT: pminuw %xmm0, %xmm1
699; SSE41-NEXT: pcmpeqw %xmm1, %xmm0
700; SSE41-NEXT: retq
701;
702; SSE42-LABEL: le_v8i16:
703; SSE42: # BB#0:
704; SSE42-NEXT: pminuw %xmm0, %xmm1
705; SSE42-NEXT: pcmpeqw %xmm1, %xmm0
706; SSE42-NEXT: retq
707;
708; AVX-LABEL: le_v8i16:
709; AVX: # BB#0:
710; AVX-NEXT: vpminuw %xmm1, %xmm0, %xmm1
711; AVX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
712; AVX-NEXT: retq
713;
714; XOP-LABEL: le_v8i16:
715; XOP: # BB#0:
716; XOP-NEXT: vpminuw %xmm1, %xmm0, %xmm1
717; XOP-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
718; XOP-NEXT: retq
719 %1 = icmp ule <8 x i16> %a, %b
720 %2 = sext <8 x i1> %1 to <8 x i16>
721 ret <8 x i16> %2
722}
723
724define <16 x i8> @le_v16i8(<16 x i8> %a, <16 x i8> %b) nounwind {
725; SSE-LABEL: le_v16i8:
726; SSE: # BB#0:
727; SSE-NEXT: pminub %xmm0, %xmm1
728; SSE-NEXT: pcmpeqb %xmm1, %xmm0
729; SSE-NEXT: retq
730;
731; AVX-LABEL: le_v16i8:
732; AVX: # BB#0:
733; AVX-NEXT: vpminub %xmm1, %xmm0, %xmm1
734; AVX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
735; AVX-NEXT: retq
736;
737; XOP-LABEL: le_v16i8:
738; XOP: # BB#0:
739; XOP-NEXT: vpminub %xmm1, %xmm0, %xmm1
740; XOP-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
741; XOP-NEXT: retq
742 %1 = icmp ule <16 x i8> %a, %b
743 %2 = sext <16 x i1> %1 to <16 x i8>
744 ret <16 x i8> %2
745}
746
747;
748; Less Than
749;
750
751define <2 x i64> @lt_v2i64(<2 x i64> %a, <2 x i64> %b) nounwind {
752; SSE2-LABEL: lt_v2i64:
753; SSE2: # BB#0:
754; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
755; SSE2-NEXT: pxor %xmm2, %xmm0
756; SSE2-NEXT: pxor %xmm2, %xmm1
757; SSE2-NEXT: movdqa %xmm1, %xmm2
758; SSE2-NEXT: pcmpgtd %xmm0, %xmm2
759; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
760; SSE2-NEXT: pcmpeqd %xmm0, %xmm1
761; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
762; SSE2-NEXT: pand %xmm3, %xmm1
763; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
764; SSE2-NEXT: por %xmm1, %xmm0
765; SSE2-NEXT: retq
766;
767; SSE41-LABEL: lt_v2i64:
768; SSE41: # BB#0:
769; SSE41-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
770; SSE41-NEXT: pxor %xmm2, %xmm0
771; SSE41-NEXT: pxor %xmm2, %xmm1
772; SSE41-NEXT: movdqa %xmm1, %xmm2
773; SSE41-NEXT: pcmpgtd %xmm0, %xmm2
774; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
775; SSE41-NEXT: pcmpeqd %xmm0, %xmm1
776; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
777; SSE41-NEXT: pand %xmm3, %xmm1
778; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
779; SSE41-NEXT: por %xmm1, %xmm0
780; SSE41-NEXT: retq
781;
782; SSE42-LABEL: lt_v2i64:
783; SSE42: # BB#0:
784; SSE42-NEXT: movdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
785; SSE42-NEXT: pxor %xmm2, %xmm0
786; SSE42-NEXT: pxor %xmm1, %xmm2
787; SSE42-NEXT: pcmpgtq %xmm0, %xmm2
788; SSE42-NEXT: movdqa %xmm2, %xmm0
789; SSE42-NEXT: retq
790;
791; AVX-LABEL: lt_v2i64:
792; AVX: # BB#0:
793; AVX-NEXT: vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
794; AVX-NEXT: vpxor %xmm2, %xmm0, %xmm0
795; AVX-NEXT: vpxor %xmm2, %xmm1, %xmm1
796; AVX-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0
797; AVX-NEXT: retq
798;
799; XOP-LABEL: lt_v2i64:
800; XOP: # BB#0:
801; XOP-NEXT: vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
802; XOP-NEXT: vpxor %xmm2, %xmm0, %xmm0
803; XOP-NEXT: vpxor %xmm2, %xmm1, %xmm1
804; XOP-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0
805; XOP-NEXT: retq
806 %1 = icmp ult <2 x i64> %a, %b
807 %2 = sext <2 x i1> %1 to <2 x i64>
808 ret <2 x i64> %2
809}
810
811define <4 x i32> @lt_v4i32(<4 x i32> %a, <4 x i32> %b) nounwind {
812; SSE-LABEL: lt_v4i32:
813; SSE: # BB#0:
814; SSE-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
815; SSE-NEXT: pxor %xmm2, %xmm0
816; SSE-NEXT: pxor %xmm1, %xmm2
817; SSE-NEXT: pcmpgtd %xmm0, %xmm2
818; SSE-NEXT: movdqa %xmm2, %xmm0
819; SSE-NEXT: retq
820;
821; AVX1-LABEL: lt_v4i32:
822; AVX1: # BB#0:
823; AVX1-NEXT: vmovdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
824; AVX1-NEXT: vpxor %xmm2, %xmm0, %xmm0
825; AVX1-NEXT: vpxor %xmm2, %xmm1, %xmm1
826; AVX1-NEXT: vpcmpgtd %xmm0, %xmm1, %xmm0
827; AVX1-NEXT: retq
828;
829; AVX2-LABEL: lt_v4i32:
830; AVX2: # BB#0:
831; AVX2-NEXT: vpbroadcastd {{.*}}(%rip), %xmm2
832; AVX2-NEXT: vpxor %xmm2, %xmm0, %xmm0
833; AVX2-NEXT: vpxor %xmm2, %xmm1, %xmm1
834; AVX2-NEXT: vpcmpgtd %xmm0, %xmm1, %xmm0
835; AVX2-NEXT: retq
836;
837; XOPAVX1-LABEL: lt_v4i32:
838; XOPAVX1: # BB#0:
839; XOPAVX1-NEXT: vmovdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
840; XOPAVX1-NEXT: vpxor %xmm2, %xmm0, %xmm0
841; XOPAVX1-NEXT: vpxor %xmm2, %xmm1, %xmm1
842; XOPAVX1-NEXT: vpcmpgtd %xmm0, %xmm1, %xmm0
843; XOPAVX1-NEXT: retq
844;
845; XOPAVX2-LABEL: lt_v4i32:
846; XOPAVX2: # BB#0:
847; XOPAVX2-NEXT: vpbroadcastd {{.*}}(%rip), %xmm2
848; XOPAVX2-NEXT: vpxor %xmm2, %xmm0, %xmm0
849; XOPAVX2-NEXT: vpxor %xmm2, %xmm1, %xmm1
850; XOPAVX2-NEXT: vpcmpgtd %xmm0, %xmm1, %xmm0
851; XOPAVX2-NEXT: retq
852;
853; AVX512-LABEL: lt_v4i32:
854; AVX512: # BB#0:
855; AVX512-NEXT: vpbroadcastd {{.*}}(%rip), %xmm2
856; AVX512-NEXT: vpxor %xmm2, %xmm0, %xmm0
857; AVX512-NEXT: vpxor %xmm2, %xmm1, %xmm1
858; AVX512-NEXT: vpcmpgtd %xmm0, %xmm1, %xmm0
859; AVX512-NEXT: retq
860 %1 = icmp ult <4 x i32> %a, %b
861 %2 = sext <4 x i1> %1 to <4 x i32>
862 ret <4 x i32> %2
863}
864
865define <8 x i16> @lt_v8i16(<8 x i16> %a, <8 x i16> %b) nounwind {
866; SSE-LABEL: lt_v8i16:
867; SSE: # BB#0:
868; SSE-NEXT: movdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
869; SSE-NEXT: pxor %xmm2, %xmm0
870; SSE-NEXT: pxor %xmm1, %xmm2
871; SSE-NEXT: pcmpgtw %xmm0, %xmm2
872; SSE-NEXT: movdqa %xmm2, %xmm0
873; SSE-NEXT: retq
874;
875; AVX-LABEL: lt_v8i16:
876; AVX: # BB#0:
877; AVX-NEXT: vmovdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
878; AVX-NEXT: vpxor %xmm2, %xmm0, %xmm0
879; AVX-NEXT: vpxor %xmm2, %xmm1, %xmm1
880; AVX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
881; AVX-NEXT: retq
882;
883; XOP-LABEL: lt_v8i16:
884; XOP: # BB#0:
885; XOP-NEXT: vmovdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
886; XOP-NEXT: vpxor %xmm2, %xmm0, %xmm0
887; XOP-NEXT: vpxor %xmm2, %xmm1, %xmm1
888; XOP-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
889; XOP-NEXT: retq
890 %1 = icmp ult <8 x i16> %a, %b
891 %2 = sext <8 x i1> %1 to <8 x i16>
892 ret <8 x i16> %2
893}
894
895define <16 x i8> @lt_v16i8(<16 x i8> %a, <16 x i8> %b) nounwind {
896; SSE-LABEL: lt_v16i8:
897; SSE: # BB#0:
898; SSE-NEXT: movdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
899; SSE-NEXT: pxor %xmm2, %xmm0
900; SSE-NEXT: pxor %xmm1, %xmm2
901; SSE-NEXT: pcmpgtb %xmm0, %xmm2
902; SSE-NEXT: movdqa %xmm2, %xmm0
903; SSE-NEXT: retq
904;
905; AVX-LABEL: lt_v16i8:
906; AVX: # BB#0:
907; AVX-NEXT: vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
908; AVX-NEXT: vpxor %xmm2, %xmm0, %xmm0
909; AVX-NEXT: vpxor %xmm2, %xmm1, %xmm1
910; AVX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
911; AVX-NEXT: retq
912;
913; XOP-LABEL: lt_v16i8:
914; XOP: # BB#0:
915; XOP-NEXT: vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
916; XOP-NEXT: vpxor %xmm2, %xmm0, %xmm0
917; XOP-NEXT: vpxor %xmm2, %xmm1, %xmm1
918; XOP-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
919; XOP-NEXT: retq
920 %1 = icmp ult <16 x i8> %a, %b
921 %2 = sext <16 x i1> %1 to <16 x i8>
922 ret <16 x i8> %2
923}