• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=SSE --check-prefix=SSE2
3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=SSE --check-prefix=SSE41
4; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefix=AVX
5
6define <16 x i8> @v16i8_icmp_uge(<16 x i8> %a, <16 x i8> %b) nounwind readnone ssp uwtable {
7; SSE-LABEL: v16i8_icmp_uge:
8; SSE:       # %bb.0:
9; SSE-NEXT:    pmaxub %xmm0, %xmm1
10; SSE-NEXT:    pcmpeqb %xmm1, %xmm0
11; SSE-NEXT:    retq
12;
13; AVX-LABEL: v16i8_icmp_uge:
14; AVX:       # %bb.0:
15; AVX-NEXT:    vpmaxub %xmm1, %xmm0, %xmm1
16; AVX-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
17; AVX-NEXT:    retq
18  %1 = icmp uge <16 x i8> %a, %b
19  %2 = sext <16 x i1> %1 to <16 x i8>
20  ret <16 x i8> %2
21}
22
23define <16 x i8> @v16i8_icmp_ule(<16 x i8> %a, <16 x i8> %b) nounwind readnone ssp uwtable {
24; SSE-LABEL: v16i8_icmp_ule:
25; SSE:       # %bb.0:
26; SSE-NEXT:    pminub %xmm0, %xmm1
27; SSE-NEXT:    pcmpeqb %xmm1, %xmm0
28; SSE-NEXT:    retq
29;
30; AVX-LABEL: v16i8_icmp_ule:
31; AVX:       # %bb.0:
32; AVX-NEXT:    vpminub %xmm1, %xmm0, %xmm1
33; AVX-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
34; AVX-NEXT:    retq
35  %1 = icmp ule <16 x i8> %a, %b
36  %2 = sext <16 x i1> %1 to <16 x i8>
37  ret <16 x i8> %2
38}
39
40define <8 x i16> @v8i16_icmp_uge(<8 x i16> %a, <8 x i16> %b) nounwind readnone ssp uwtable {
41; SSE2-LABEL: v8i16_icmp_uge:
42; SSE2:       # %bb.0:
43; SSE2-NEXT:    psubusw %xmm0, %xmm1
44; SSE2-NEXT:    pxor %xmm0, %xmm0
45; SSE2-NEXT:    pcmpeqw %xmm1, %xmm0
46; SSE2-NEXT:    retq
47;
48; SSE41-LABEL: v8i16_icmp_uge:
49; SSE41:       # %bb.0:
50; SSE41-NEXT:    pmaxuw %xmm0, %xmm1
51; SSE41-NEXT:    pcmpeqw %xmm1, %xmm0
52; SSE41-NEXT:    retq
53;
54; AVX-LABEL: v8i16_icmp_uge:
55; AVX:       # %bb.0:
56; AVX-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm1
57; AVX-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
58; AVX-NEXT:    retq
59  %1 = icmp uge <8 x i16> %a, %b
60  %2 = sext <8 x i1> %1 to <8 x i16>
61  ret <8 x i16> %2
62}
63
64define <8 x i16> @v8i16_icmp_ule(<8 x i16> %a, <8 x i16> %b) nounwind readnone ssp uwtable {
65; SSE2-LABEL: v8i16_icmp_ule:
66; SSE2:       # %bb.0:
67; SSE2-NEXT:    psubusw %xmm1, %xmm0
68; SSE2-NEXT:    pxor %xmm1, %xmm1
69; SSE2-NEXT:    pcmpeqw %xmm1, %xmm0
70; SSE2-NEXT:    retq
71;
72; SSE41-LABEL: v8i16_icmp_ule:
73; SSE41:       # %bb.0:
74; SSE41-NEXT:    pminuw %xmm0, %xmm1
75; SSE41-NEXT:    pcmpeqw %xmm1, %xmm0
76; SSE41-NEXT:    retq
77;
78; AVX-LABEL: v8i16_icmp_ule:
79; AVX:       # %bb.0:
80; AVX-NEXT:    vpminuw %xmm1, %xmm0, %xmm1
81; AVX-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
82; AVX-NEXT:    retq
83  %1 = icmp ule <8 x i16> %a, %b
84  %2 = sext <8 x i1> %1 to <8 x i16>
85  ret <8 x i16> %2
86}
87
88define <4 x i32> @v4i32_icmp_uge(<4 x i32> %a, <4 x i32> %b) nounwind readnone ssp uwtable {
89; SSE2-LABEL: v4i32_icmp_uge:
90; SSE2:       # %bb.0:
91; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
92; SSE2-NEXT:    pxor %xmm2, %xmm0
93; SSE2-NEXT:    pxor %xmm1, %xmm2
94; SSE2-NEXT:    pcmpgtd %xmm0, %xmm2
95; SSE2-NEXT:    pcmpeqd %xmm0, %xmm0
96; SSE2-NEXT:    pxor %xmm2, %xmm0
97; SSE2-NEXT:    retq
98;
99; SSE41-LABEL: v4i32_icmp_uge:
100; SSE41:       # %bb.0:
101; SSE41-NEXT:    pmaxud %xmm0, %xmm1
102; SSE41-NEXT:    pcmpeqd %xmm1, %xmm0
103; SSE41-NEXT:    retq
104;
105; AVX-LABEL: v4i32_icmp_uge:
106; AVX:       # %bb.0:
107; AVX-NEXT:    vpmaxud %xmm1, %xmm0, %xmm1
108; AVX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
109; AVX-NEXT:    retq
110  %1 = icmp uge <4 x i32> %a, %b
111  %2 = sext <4 x i1> %1 to <4 x i32>
112  ret <4 x i32> %2
113}
114
115define <4 x i32> @v4i32_icmp_ule(<4 x i32> %a, <4 x i32> %b) nounwind readnone ssp uwtable {
116; SSE2-LABEL: v4i32_icmp_ule:
117; SSE2:       # %bb.0:
118; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
119; SSE2-NEXT:    pxor %xmm2, %xmm1
120; SSE2-NEXT:    pxor %xmm2, %xmm0
121; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
122; SSE2-NEXT:    pcmpeqd %xmm1, %xmm1
123; SSE2-NEXT:    pxor %xmm1, %xmm0
124; SSE2-NEXT:    retq
125;
126; SSE41-LABEL: v4i32_icmp_ule:
127; SSE41:       # %bb.0:
128; SSE41-NEXT:    pminud %xmm0, %xmm1
129; SSE41-NEXT:    pcmpeqd %xmm1, %xmm0
130; SSE41-NEXT:    retq
131;
132; AVX-LABEL: v4i32_icmp_ule:
133; AVX:       # %bb.0:
134; AVX-NEXT:    vpminud %xmm1, %xmm0, %xmm1
135; AVX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
136; AVX-NEXT:    retq
137  %1 = icmp ule <4 x i32> %a, %b
138  %2 = sext <4 x i1> %1 to <4 x i32>
139  ret <4 x i32> %2
140}
141
142; At one point we were incorrectly constant-folding a setcc to 0x1 instead of
143; 0xff, leading to a constpool load. The instruction doesn't matter here, but it
144; should set all bits to 1.
145define <16 x i8> @test_setcc_constfold_vi8(<16 x i8> %l, <16 x i8> %r) {
146; SSE-LABEL: test_setcc_constfold_vi8:
147; SSE:       # %bb.0:
148; SSE-NEXT:    pcmpeqd %xmm0, %xmm0
149; SSE-NEXT:    retq
150;
151; AVX-LABEL: test_setcc_constfold_vi8:
152; AVX:       # %bb.0:
153; AVX-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
154; AVX-NEXT:    retq
155  %test1 = icmp eq <16 x i8> %l, %r
156  %mask1 = sext <16 x i1> %test1 to <16 x i8>
157  %test2 = icmp ne <16 x i8> %l, %r
158  %mask2 = sext <16 x i1> %test2 to <16 x i8>
159  %res = or <16 x i8> %mask1, %mask2
160  ret <16 x i8> %res
161}
162
163; Make sure sensible results come from doing extension afterwards
164define <16 x i8> @test_setcc_constfold_vi1(<16 x i8> %l, <16 x i8> %r) {
165; SSE-LABEL: test_setcc_constfold_vi1:
166; SSE:       # %bb.0:
167; SSE-NEXT:    pcmpeqd %xmm0, %xmm0
168; SSE-NEXT:    retq
169;
170; AVX-LABEL: test_setcc_constfold_vi1:
171; AVX:       # %bb.0:
172; AVX-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
173; AVX-NEXT:    retq
174  %test1 = icmp eq <16 x i8> %l, %r
175  %test2 = icmp ne <16 x i8> %l, %r
176  %res = or <16 x i1> %test1, %test2
177  %mask = sext <16 x i1> %res to <16 x i8>
178  ret <16 x i8> %mask
179}
180
181; 64-bit case is also particularly important, as the constant "-1" is probably
182; just 32-bits wide.
183define <2 x i64> @test_setcc_constfold_vi64(<2 x i64> %l, <2 x i64> %r) {
184; SSE-LABEL: test_setcc_constfold_vi64:
185; SSE:       # %bb.0:
186; SSE-NEXT:    pcmpeqd %xmm0, %xmm0
187; SSE-NEXT:    retq
188;
189; AVX-LABEL: test_setcc_constfold_vi64:
190; AVX:       # %bb.0:
191; AVX-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
192; AVX-NEXT:    retq
193  %test1 = icmp eq <2 x i64> %l, %r
194  %mask1 = sext <2 x i1> %test1 to <2 x i64>
195  %test2 = icmp ne <2 x i64> %l, %r
196  %mask2 = sext <2 x i1> %test2 to <2 x i64>
197  %res = or <2 x i64> %mask1, %mask2
198  ret <2 x i64> %res
199}
200