| blob | bc0dd022bfae400e66de08cd64965a01beafb537 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=avx512fp16,avx512vl -O3 | FileCheck %s --check-prefixes=CHECK
3 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512fp16,avx512vl -O3 | FileCheck %s --check-prefixes=CHECK
6 declare <4 x i64> @llvm.experimental.constrained.fptosi.v4i64.v4f16(<4 x half>, metadata)
7 declare <4 x i64> @llvm.experimental.constrained.fptoui.v4i64.v4f16(<4 x half>, metadata)
8 declare <8 x i32> @llvm.experimental.constrained.fptosi.v8i32.v8f16(<8 x half>, metadata)
9 declare <8 x i32> @llvm.experimental.constrained.fptoui.v8i32.v8f16(<8 x half>, metadata)
10 declare <16 x i16> @llvm.experimental.constrained.fptosi.v16i16.v16f16(<16 x half>, metadata)
11 declare <16 x i16> @llvm.experimental.constrained.fptoui.v16i16.v16f16(<16 x half>, metadata)
12 declare <16 x i8> @llvm.experimental.constrained.fptosi.v16i8.v16f16(<16 x half>, metadata)
13 declare <16 x i8> @llvm.experimental.constrained.fptoui.v16i8.v16f16(<16 x half>, metadata)
14 declare <16 x i1> @llvm.experimental.constrained.fptosi.v16i1.v16f16(<16 x half>, metadata)
15 declare <16 x i1> @llvm.experimental.constrained.fptoui.v16i1.v16f16(<16 x half>, metadata)
17 define <4 x i64> @strict_vector_fptosi_v4f16_to_v4i64(<4 x half> %a) #0 {
18 ; CHECK-LABEL: strict_vector_fptosi_v4f16_to_v4i64:
19 ; CHECK: # %bb.0:
20 ; CHECK-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero
21 ; CHECK-NEXT: vcvttph2qq %xmm0, %ymm0
22 ; CHECK-NEXT: ret{{[l|q]}}
23 %ret = call <4 x i64> @llvm.experimental.constrained.fptosi.v4i64.v4f16(<4 x half> %a,
24 metadata !"fpexcept.strict") #0
25 ret <4 x i64> %ret
26 }
28 define <4 x i64> @strict_vector_fptoui_v4f16_to_v4i64(<4 x half> %a) #0 {
29 ; CHECK-LABEL: strict_vector_fptoui_v4f16_to_v4i64:
30 ; CHECK: # %bb.0:
31 ; CHECK-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero
32 ; CHECK-NEXT: vcvttph2uqq %xmm0, %ymm0
33 ; CHECK-NEXT: ret{{[l|q]}}
34 %ret = call <4 x i64> @llvm.experimental.constrained.fptoui.v4i64.v4f16(<4 x half> %a,
35 metadata !"fpexcept.strict") #0
36 ret <4 x i64> %ret
37 }
39 define <8 x i32> @strict_vector_fptosi_v8f16_to_v8i32(<8 x half> %a) #0 {
40 ; CHECK-LABEL: strict_vector_fptosi_v8f16_to_v8i32:
41 ; CHECK: # %bb.0:
42 ; CHECK-NEXT: vcvttph2dq %xmm0, %ymm0
43 ; CHECK-NEXT: ret{{[l|q]}}
44 %ret = call <8 x i32> @llvm.experimental.constrained.fptosi.v8i32.v8f16(<8 x half> %a,
45 metadata !"fpexcept.strict") #0
46 ret <8 x i32> %ret
47 }
49 define <8 x i32> @strict_vector_fptoui_v8f16_to_v8i32(<8 x half> %a) #0 {
50 ; CHECK-LABEL: strict_vector_fptoui_v8f16_to_v8i32:
51 ; CHECK: # %bb.0:
52 ; CHECK-NEXT: vcvttph2udq %xmm0, %ymm0
53 ; CHECK-NEXT: ret{{[l|q]}}
54 %ret = call <8 x i32> @llvm.experimental.constrained.fptoui.v8i32.v8f16(<8 x half> %a,
55 metadata !"fpexcept.strict") #0
56 ret <8 x i32> %ret
57 }
59 define <16 x i16> @strict_vector_fptosi_v16f16_to_v16i16(<16 x half> %a) #0 {
60 ; CHECK-LABEL: strict_vector_fptosi_v16f16_to_v16i16:
61 ; CHECK: # %bb.0:
62 ; CHECK-NEXT: vcvttph2dq %ymm0, %zmm0
63 ; CHECK-NEXT: vpmovdw %zmm0, %ymm0
64 ; CHECK-NEXT: ret{{[l|q]}}
65 %ret = call <16 x i16> @llvm.experimental.constrained.fptosi.v16i16.v16f16(<16 x half> %a,
66 metadata !"fpexcept.strict") #0
67 ret <16 x i16> %ret
68 }
70 define <16 x i16> @strict_vector_fptoui_v16f16_to_v16i16(<16 x half> %a) #0 {
71 ; CHECK-LABEL: strict_vector_fptoui_v16f16_to_v16i16:
72 ; CHECK: # %bb.0:
73 ; CHECK-NEXT: vcvttph2dq %ymm0, %zmm0
74 ; CHECK-NEXT: vpmovdw %zmm0, %ymm0
75 ; CHECK-NEXT: ret{{[l|q]}}
76 %ret = call <16 x i16> @llvm.experimental.constrained.fptoui.v16i16.v16f16(<16 x half> %a,
77 metadata !"fpexcept.strict") #0
78 ret <16 x i16> %ret
79 }
81 define <16 x i8> @strict_vector_fptosi_v16f16_to_v16i8(<16 x half> %a) #0 {
82 ; CHECK-LABEL: strict_vector_fptosi_v16f16_to_v16i8:
83 ; CHECK: # %bb.0:
84 ; CHECK-NEXT: vcvttph2dq %ymm0, %zmm0
85 ; CHECK-NEXT: vpmovdb %zmm0, %xmm0
86 ; CHECK-NEXT: vzeroupper
87 ; CHECK-NEXT: ret{{[l|q]}}
88 %ret = call <16 x i8> @llvm.experimental.constrained.fptosi.v16i8.v16f16(<16 x half> %a,
89 metadata !"fpexcept.strict") #0
90 ret <16 x i8> %ret
91 }
93 define <16 x i8> @strict_vector_fptoui_v16f16_to_v16i8(<16 x half> %a) #0 {
94 ; CHECK-LABEL: strict_vector_fptoui_v16f16_to_v16i8:
95 ; CHECK: # %bb.0:
96 ; CHECK-NEXT: vcvttph2dq %ymm0, %zmm0
97 ; CHECK-NEXT: vpmovdb %zmm0, %xmm0
98 ; CHECK-NEXT: vzeroupper
99 ; CHECK-NEXT: ret{{[l|q]}}
100 %ret = call <16 x i8> @llvm.experimental.constrained.fptoui.v16i8.v16f16(<16 x half> %a,
101 metadata !"fpexcept.strict") #0
102 ret <16 x i8> %ret
103 }
105 define <16 x i1> @strict_vector_fptosi_v16f16_to_v16i1(<16 x half> %a) #0 {
106 ; CHECK-LABEL: strict_vector_fptosi_v16f16_to_v16i1:
107 ; CHECK: # %bb.0:
108 ; CHECK-NEXT: vcvttph2dq %ymm0, %zmm0
109 ; CHECK-NEXT: vpmovd2m %zmm0, %k0
110 ; CHECK-NEXT: vpmovm2b %k0, %xmm0
111 ; CHECK-NEXT: vzeroupper
112 ; CHECK-NEXT: ret{{[l|q]}}
113 %ret = call <16 x i1> @llvm.experimental.constrained.fptosi.v16i1.v16f16(<16 x half> %a,
114 metadata !"fpexcept.strict") #0
115 ret <16 x i1> %ret
116 }
118 define <16 x i1> @strict_vector_fptoui_v16f16_to_v16i1(<16 x half> %a) #0 {
119 ; CHECK-LABEL: strict_vector_fptoui_v16f16_to_v16i1:
120 ; CHECK: # %bb.0:
121 ; CHECK-NEXT: vcvttph2dq %ymm0, %zmm0
122 ; CHECK-NEXT: vpslld $31, %zmm0, %zmm0
123 ; CHECK-NEXT: vpmovd2m %zmm0, %k0
124 ; CHECK-NEXT: vpmovm2b %k0, %xmm0
125 ; CHECK-NEXT: vzeroupper
126 ; CHECK-NEXT: ret{{[l|q]}}
127 %ret = call <16 x i1> @llvm.experimental.constrained.fptoui.v16i1.v16f16(<16 x half> %a,
128 metadata !"fpexcept.strict") #0
129 ret <16 x i1> %ret
130 }
132 attributes #0 = { strictfp }