• 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-apple-macosx -mattr=+avx | FileCheck %s --check-prefix=AVX --check-prefix=AVX1
3; RUN: llc < %s -mtriple=x86_64-apple-macosx -mattr=+avx2 | FileCheck %s --check-prefix=AVX --check-prefix=AVX2
4
5target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
6
7; For this test we used to optimize the <i1 true, i1 false, i1 false, i1 true>
8; mask into <i32 2147483648, i32 0, i32 0, i32 2147483648> because we thought
9; we would lower that into a blend where only the high bit is relevant.
10; However, since the whole mask is constant, this is simplified incorrectly
11; by the generic code, because it was expecting -1 in place of 2147483648.
12;
13; The problem does not occur without AVX, because vselect of v4i32 is not legal
14; nor custom.
15;
16; <rdar://problem/18675020>
17
18define void @test(<4 x i16>* %a, <4 x i16>* %b) {
19; AVX-LABEL: test:
20; AVX:       ## %bb.0: ## %body
21; AVX-NEXT:    movq {{.*}}(%rip), %rax
22; AVX-NEXT:    movq %rax, (%rdi)
23; AVX-NEXT:    movq {{.*}}(%rip), %rax
24; AVX-NEXT:    movq %rax, (%rsi)
25; AVX-NEXT:    retq
26body:
27  %predphi = select <4 x i1> <i1 true, i1 false, i1 false, i1 true>, <4 x i16> <i16 -3, i16 545, i16 4385, i16 14807>, <4 x i16> <i16 123, i16 124, i16 125, i16 127>
28  %predphi42 = select <4 x i1> <i1 true, i1 false, i1 false, i1 true>, <4 x i16> <i16 -1, i16 -1, i16 -1, i16 -1>, <4 x i16> zeroinitializer
29  store <4 x i16> %predphi, <4 x i16>* %a, align 8
30  store <4 x i16> %predphi42, <4 x i16>* %b, align 8
31  ret void
32}
33
34; Improve code coverage.
35;
36; When shrinking the condition used into the select to match a blend, this
37; test case exercises the path where the modified node is not the root
38; of the condition.
39
40define void @test2(double** %call1559, i64 %indvars.iv4198, <4 x i1> %tmp1895) {
41; AVX1-LABEL: test2:
42; AVX1:       ## %bb.0: ## %bb
43; AVX1-NEXT:    vpslld $31, %xmm0, %xmm0
44; AVX1-NEXT:    vpsrad $31, %xmm0, %xmm0
45; AVX1-NEXT:    vpmovsxdq %xmm0, %xmm1
46; AVX1-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
47; AVX1-NEXT:    vpmovsxdq %xmm0, %xmm0
48; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
49; AVX1-NEXT:    movq (%rdi,%rsi,8), %rax
50; AVX1-NEXT:    vmovapd {{.*#+}} ymm1 = [5.000000e-01,5.000000e-01,5.000000e-01,5.000000e-01]
51; AVX1-NEXT:    vblendvpd %ymm0, {{.*}}(%rip), %ymm1, %ymm0
52; AVX1-NEXT:    vmovupd %ymm0, (%rax)
53; AVX1-NEXT:    vzeroupper
54; AVX1-NEXT:    retq
55;
56; AVX2-LABEL: test2:
57; AVX2:       ## %bb.0: ## %bb
58; AVX2-NEXT:    vpslld $31, %xmm0, %xmm0
59; AVX2-NEXT:    vpmovsxdq %xmm0, %ymm0
60; AVX2-NEXT:    movq (%rdi,%rsi,8), %rax
61; AVX2-NEXT:    vbroadcastsd {{.*#+}} ymm1 = [-0.5,-0.5,-0.5,-0.5]
62; AVX2-NEXT:    vbroadcastsd {{.*#+}} ymm2 = [0.5,0.5,0.5,0.5]
63; AVX2-NEXT:    vblendvpd %ymm0, %ymm1, %ymm2, %ymm0
64; AVX2-NEXT:    vmovupd %ymm0, (%rax)
65; AVX2-NEXT:    vzeroupper
66; AVX2-NEXT:    retq
67bb:
68  %arrayidx1928 = getelementptr inbounds double*, double** %call1559, i64 %indvars.iv4198
69  %tmp1888 = load double*, double** %arrayidx1928, align 8
70  %predphi.v.v = select <4 x i1> %tmp1895, <4 x double> <double -5.000000e-01, double -5.000000e-01, double -5.000000e-01, double -5.000000e-01>, <4 x double> <double 5.000000e-01, double 5.000000e-01, double 5.000000e-01, double 5.000000e-01>
71  %tmp1900 = bitcast double* %tmp1888 to <4 x double>*
72  store <4 x double> %predphi.v.v, <4 x double>* %tmp1900, align 8
73  ret void
74}
75
76; For this test, we used to optimized the conditional mask for the blend, i.e.,
77; we shrunk some of its bits.
78; However, this same mask was used in another select (%predphi31) that turned out
79; to be optimized into a and. In that case, the conditional mask was wrong.
80;
81; Make sure that the and is fed by the original mask.
82;
83; <rdar://problem/18819506>
84
85define void @test3(<4 x i32> %induction30, <4 x i16>* %tmp16, <4 x i16>* %tmp17,  <4 x i16> %tmp3, <4 x i16> %tmp12) {
86; AVX1-LABEL: test3:
87; AVX1:       ## %bb.0:
88; AVX1-NEXT:    vpshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
89; AVX1-NEXT:    vmovdqa {{.*#+}} xmm4 = [1431655766,1431655766,1431655766,1431655766]
90; AVX1-NEXT:    vpmuldq %xmm4, %xmm3, %xmm3
91; AVX1-NEXT:    vpmuldq %xmm4, %xmm0, %xmm4
92; AVX1-NEXT:    vpshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
93; AVX1-NEXT:    vpblendw {{.*#+}} xmm3 = xmm4[0,1],xmm3[2,3],xmm4[4,5],xmm3[6,7]
94; AVX1-NEXT:    vpsrld $31, %xmm3, %xmm4
95; AVX1-NEXT:    vpaddd %xmm4, %xmm3, %xmm3
96; AVX1-NEXT:    vpmulld {{.*}}(%rip), %xmm3, %xmm3
97; AVX1-NEXT:    vpsubd %xmm3, %xmm0, %xmm0
98; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
99; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm0, %xmm0
100; AVX1-NEXT:    vblendvps %xmm0, %xmm1, %xmm2, %xmm1
101; AVX1-NEXT:    vpackssdw %xmm0, %xmm0, %xmm0
102; AVX1-NEXT:    vmovq %xmm0, (%rdi)
103; AVX1-NEXT:    vpshufb {{.*#+}} xmm0 = xmm1[0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15]
104; AVX1-NEXT:    vmovq %xmm0, (%rsi)
105; AVX1-NEXT:    retq
106;
107; AVX2-LABEL: test3:
108; AVX2:       ## %bb.0:
109; AVX2-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [1431655766,1431655766,1431655766,1431655766]
110; AVX2-NEXT:    vpshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
111; AVX2-NEXT:    vpshufd {{.*#+}} xmm5 = xmm0[1,1,3,3]
112; AVX2-NEXT:    vpmuldq %xmm4, %xmm5, %xmm4
113; AVX2-NEXT:    vpmuldq %xmm3, %xmm0, %xmm3
114; AVX2-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
115; AVX2-NEXT:    vpblendd {{.*#+}} xmm3 = xmm3[0],xmm4[1],xmm3[2],xmm4[3]
116; AVX2-NEXT:    vpsrld $31, %xmm3, %xmm4
117; AVX2-NEXT:    vpaddd %xmm4, %xmm3, %xmm3
118; AVX2-NEXT:    vpbroadcastd {{.*#+}} xmm4 = [3,3,3,3]
119; AVX2-NEXT:    vpmulld %xmm4, %xmm3, %xmm3
120; AVX2-NEXT:    vpsubd %xmm3, %xmm0, %xmm0
121; AVX2-NEXT:    vpxor %xmm3, %xmm3, %xmm3
122; AVX2-NEXT:    vpcmpeqd %xmm3, %xmm0, %xmm0
123; AVX2-NEXT:    vblendvps %xmm0, %xmm1, %xmm2, %xmm1
124; AVX2-NEXT:    vpackssdw %xmm0, %xmm0, %xmm0
125; AVX2-NEXT:    vmovq %xmm0, (%rdi)
126; AVX2-NEXT:    vpshufb {{.*#+}} xmm0 = xmm1[0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15]
127; AVX2-NEXT:    vmovq %xmm0, (%rsi)
128; AVX2-NEXT:    retq
129  %tmp6 = srem <4 x i32> %induction30, <i32 3, i32 3, i32 3, i32 3>
130  %tmp7 = icmp eq <4 x i32> %tmp6, zeroinitializer
131  %predphi = select <4 x i1> %tmp7, <4 x i16> %tmp3, <4 x i16> %tmp12
132  %predphi31 = select <4 x i1> %tmp7, <4 x i16> <i16 -1, i16 -1, i16 -1, i16 -1>, <4 x i16> zeroinitializer
133
134  store <4 x i16> %predphi31, <4 x i16>* %tmp16, align 8
135  store <4 x i16> %predphi, <4 x i16>* %tmp17, align 8
136 ret void
137}
138
139; We shouldn't try to lower this directly using VSELECT because we don't have
140; vpblendvb in AVX1, only in AVX2. Instead, it should be expanded.
141
142define <32 x i8> @PR22706(<32 x i1> %x) {
143; AVX1-LABEL: PR22706:
144; AVX1:       ## %bb.0:
145; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
146; AVX1-NEXT:    vpsllw $7, %xmm1, %xmm1
147; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
148; AVX1-NEXT:    vpand %xmm2, %xmm1, %xmm1
149; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
150; AVX1-NEXT:    vpcmpgtb %xmm1, %xmm3, %xmm1
151; AVX1-NEXT:    vmovdqa {{.*#+}} xmm4 = [2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2]
152; AVX1-NEXT:    vpaddb %xmm4, %xmm1, %xmm1
153; AVX1-NEXT:    vpsllw $7, %xmm0, %xmm0
154; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
155; AVX1-NEXT:    vpcmpgtb %xmm0, %xmm3, %xmm0
156; AVX1-NEXT:    vpaddb %xmm4, %xmm0, %xmm0
157; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
158; AVX1-NEXT:    retq
159;
160; AVX2-LABEL: PR22706:
161; AVX2:       ## %bb.0:
162; AVX2-NEXT:    vpsllw $7, %ymm0, %ymm0
163; AVX2-NEXT:    vpand {{.*}}(%rip), %ymm0, %ymm0
164; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
165; AVX2-NEXT:    vpcmpgtb %ymm0, %ymm1, %ymm0
166; AVX2-NEXT:    vpaddb {{.*}}(%rip), %ymm0, %ymm0
167; AVX2-NEXT:    retq
168  %tmp = select <32 x i1> %x, <32 x i8> <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>, <32 x i8> <i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2>
169  ret <32 x i8> %tmp
170}
171