1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s
3
4; (zext(select c, load1, load2)) -> (select c, zextload1, zextload2)
5define i64 @zext_scalar(i8* %p, i1 zeroext %c) {
6; CHECK-LABEL: zext_scalar:
7; CHECK:       # %bb.0:
8; CHECK-NEXT:    movzbl (%rdi), %ecx
9; CHECK-NEXT:    movzbl 1(%rdi), %eax
10; CHECK-NEXT:    testl %esi, %esi
11; CHECK-NEXT:    cmoveq %rcx, %rax
12; CHECK-NEXT:    retq
13  %ld1 = load volatile i8, i8* %p
14  %arrayidx1 = getelementptr inbounds i8, i8* %p, i64 1
15  %ld2 = load volatile i8, i8* %arrayidx1
16  %cond.v = select i1 %c, i8 %ld2, i8 %ld1
17  %cond = zext i8 %cond.v to i64
18  ret i64 %cond
19}
20
21define i64 @zext_scalar2(i8* %p, i16* %q, i1 zeroext %c) {
22; CHECK-LABEL: zext_scalar2:
23; CHECK:       # %bb.0:
24; CHECK-NEXT:    movzbl (%rdi), %ecx
25; CHECK-NEXT:    movzwl (%rsi), %eax
26; CHECK-NEXT:    testl %edx, %edx
27; CHECK-NEXT:    cmoveq %rcx, %rax
28; CHECK-NEXT:    retq
29  %ld1 = load volatile i8, i8* %p
30  %ext_ld1 = zext i8 %ld1 to i16
31  %ld2 = load volatile i16, i16* %q
32  %cond.v = select i1 %c, i16 %ld2, i16 %ext_ld1
33  %cond = zext i16 %cond.v to i64
34  ret i64 %cond
35}
36
37; Don't fold the ext if there is a load with conflicting ext type.
38define i64 @zext_scalar_neg(i8* %p, i16* %q, i1 zeroext %c) {
39; CHECK-LABEL: zext_scalar_neg:
40; CHECK:       # %bb.0:
41; CHECK-NEXT:    movsbl (%rdi), %eax
42; CHECK-NEXT:    testl %edx, %edx
43; CHECK-NEXT:    je .LBB2_2
44; CHECK-NEXT:  # %bb.1:
45; CHECK-NEXT:    movzwl (%rsi), %eax
46; CHECK-NEXT:  .LBB2_2:
47; CHECK-NEXT:    movzwl %ax, %eax
48; CHECK-NEXT:    retq
49  %ld1 = load volatile i8, i8* %p
50  %ext_ld1 = sext i8 %ld1 to i16
51  %ld2 = load volatile i16, i16* %q
52  %cond.v = select i1 %c, i16 %ld2, i16 %ext_ld1
53  %cond = zext i16 %cond.v to i64
54  ret i64 %cond
55}
56
57; (sext(select c, load1, load2)) -> (select c, sextload1, sextload2)
58define i64 @sext_scalar(i8* %p, i1 zeroext %c) {
59; CHECK-LABEL: sext_scalar:
60; CHECK:       # %bb.0:
61; CHECK-NEXT:    movsbq (%rdi), %rcx
62; CHECK-NEXT:    movsbq 1(%rdi), %rax
63; CHECK-NEXT:    testl %esi, %esi
64; CHECK-NEXT:    cmoveq %rcx, %rax
65; CHECK-NEXT:    retq
66  %ld1 = load volatile i8, i8* %p
67  %arrayidx1 = getelementptr inbounds i8, i8* %p, i64 1
68  %ld2 = load volatile i8, i8* %arrayidx1
69  %cond.v = select i1 %c, i8 %ld2, i8 %ld1
70  %cond = sext i8 %cond.v to i64
71  ret i64 %cond
72}
73
74; Same as zext_scalar, but operate on vectors.
75define <2 x i64> @zext_vector_i1(<2 x i32>* %p, i1 zeroext %c) {
76; CHECK-LABEL: zext_vector_i1:
77; CHECK:       # %bb.0:
78; CHECK-NEXT:    pmovzxdq {{.*#+}} xmm1 = mem[0],zero,mem[1],zero
79; CHECK-NEXT:    pmovzxdq {{.*#+}} xmm0 = mem[0],zero,mem[1],zero
80; CHECK-NEXT:    testl %esi, %esi
81; CHECK-NEXT:    jne .LBB4_2
82; CHECK-NEXT:  # %bb.1:
83; CHECK-NEXT:    movdqa %xmm1, %xmm0
84; CHECK-NEXT:  .LBB4_2:
85; CHECK-NEXT:    retq
86  %ld1 = load volatile <2 x i32>, <2 x i32>* %p
87  %arrayidx1 = getelementptr inbounds <2 x i32>, <2 x i32>* %p, i64 1
88  %ld2 = load volatile <2 x i32>, <2 x i32>* %arrayidx1
89  %cond.v = select i1 %c, <2 x i32> %ld2, <2 x i32> %ld1
90  %cond = zext <2 x i32> %cond.v to <2 x i64>
91  ret <2 x i64> %cond
92}
93
94define <2 x i64> @zext_vector_v2i1(<2 x i32>* %p, <2 x i1> %c) {
95; CHECK-LABEL: zext_vector_v2i1:
96; CHECK:       # %bb.0:
97; CHECK-NEXT:    psllq $63, %xmm0
98; CHECK-NEXT:    pmovzxdq {{.*#+}} xmm1 = mem[0],zero,mem[1],zero
99; CHECK-NEXT:    pmovzxdq {{.*#+}} xmm2 = mem[0],zero,mem[1],zero
100; CHECK-NEXT:    blendvpd %xmm0, %xmm2, %xmm1
101; CHECK-NEXT:    movapd %xmm1, %xmm0
102; CHECK-NEXT:    retq
103  %ld1 = load volatile <2 x i32>, <2 x i32>* %p
104  %arrayidx1 = getelementptr inbounds <2 x i32>, <2 x i32>* %p, i64 1
105  %ld2 = load volatile <2 x i32>, <2 x i32>* %arrayidx1
106  %cond.v = select <2 x i1> %c, <2 x i32> %ld2, <2 x i32> %ld1
107  %cond = zext <2 x i32> %cond.v to <2 x i64>
108  ret <2 x i64> %cond
109}
110
111; Same as sext_scalar, but operate on vectors.
112define <2 x i64> @sext_vector_i1(<2 x i32>* %p, i1 zeroext %c) {
113; CHECK-LABEL: sext_vector_i1:
114; CHECK:       # %bb.0:
115; CHECK-NEXT:    pmovsxdq (%rdi), %xmm1
116; CHECK-NEXT:    pmovsxdq 8(%rdi), %xmm0
117; CHECK-NEXT:    testl %esi, %esi
118; CHECK-NEXT:    jne .LBB6_2
119; CHECK-NEXT:  # %bb.1:
120; CHECK-NEXT:    movdqa %xmm1, %xmm0
121; CHECK-NEXT:  .LBB6_2:
122; CHECK-NEXT:    retq
123  %ld1 = load volatile <2 x i32>, <2 x i32>* %p
124  %arrayidx1 = getelementptr inbounds <2 x i32>, <2 x i32>* %p, i64 1
125  %ld2 = load volatile <2 x i32>, <2 x i32>* %arrayidx1
126  %cond.v = select i1 %c, <2 x i32> %ld2, <2 x i32> %ld1
127  %cond = sext <2 x i32> %cond.v to <2 x i64>
128  ret <2 x i64> %cond
129}
130
131define <2 x i64> @sext_vector_v2i1(<2 x i32>* %p, <2 x i1> %c) {
132; CHECK-LABEL: sext_vector_v2i1:
133; CHECK:       # %bb.0:
134; CHECK-NEXT:    psllq $63, %xmm0
135; CHECK-NEXT:    pmovsxdq (%rdi), %xmm1
136; CHECK-NEXT:    pmovsxdq 8(%rdi), %xmm2
137; CHECK-NEXT:    blendvpd %xmm0, %xmm2, %xmm1
138; CHECK-NEXT:    movapd %xmm1, %xmm0
139; CHECK-NEXT:    retq
140  %ld1 = load volatile <2 x i32>, <2 x i32>* %p
141  %arrayidx1 = getelementptr inbounds <2 x i32>, <2 x i32>* %p, i64 1
142  %ld2 = load volatile <2 x i32>, <2 x i32>* %arrayidx1
143  %cond.v = select <2 x i1> %c, <2 x i32> %ld2, <2 x i32> %ld1
144  %cond = sext <2 x i32> %cond.v to <2 x i64>
145  ret <2 x i64> %cond
146}
147