1; RUN: llc < %s -march=x86-64 -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 -mattr=-f16c | FileCheck %s -check-prefix=CHECK -check-prefix=LIBCALL
2; RUN: llc < %s -march=x86-64 -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 -mattr=+f16c | FileCheck %s -check-prefix=CHECK -check-prefix=F16C
3; RUN: llc < %s -march=x86-64 -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 -mattr=-f16c,+soft-float | FileCheck %s -check-prefix=CHECK -check-prefix=SOFTFLOAT
4; RUN: llc < %s -march=x86-64 -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 -mattr=+f16c,+soft-float | FileCheck %s -check-prefix=CHECK -check-prefix=SOFTFLOAT
5
6; This is a test for float to half float conversions on x86-64.
7;
8; If flag -soft-float is set, or if there is no F16C support, then:
9; 1) half float to float conversions are
10;    translated into calls to __gnu_h2f_ieee defined
11;    by the compiler runtime library;
12; 2) float to half float conversions are translated into calls
13;    to __gnu_f2h_ieee which expected to be defined by the
14;    compiler runtime library.
15;
16; Otherwise (we have F16C support):
17; 1) half float to float conversion are translated using
18;    vcvtph2ps instructions;
19; 2) float to half float conversions are translated using
20;    vcvtps2ph instructions
21
22
23define void @test1(float %src, i16* %dest) {
24  %1 = tail call i16 @llvm.convert.to.fp16.f32(float %src)
25  store i16 %1, i16* %dest, align 2
26  ret void
27}
28; CHECK-LABEL: test1
29; LIBCALL: callq  __gnu_f2h_ieee
30; SOFTFLOAT: callq  __gnu_f2h_ieee
31; F16C: vcvtps2ph
32; CHECK: ret
33
34
35define float @test2(i16* nocapture %src) {
36  %1 = load i16, i16* %src, align 2
37  %2 = tail call float @llvm.convert.from.fp16.f32(i16 %1)
38  ret float %2
39}
40; CHECK-LABEL: test2:
41; LIBCALL: jmp  __gnu_h2f_ieee
42; SOFTFLOAT: callq  __gnu_h2f_ieee
43; F16C: vcvtph2ps
44; F16C: ret
45
46
47define float @test3(float %src) nounwind uwtable readnone {
48  %1 = tail call i16 @llvm.convert.to.fp16.f32(float %src)
49  %2 = tail call float @llvm.convert.from.fp16.f32(i16 %1)
50  ret float %2
51}
52
53; CHECK-LABEL: test3:
54; LIBCALL: callq  __gnu_f2h_ieee
55; LIBCALL: jmp   __gnu_h2f_ieee
56; SOFTFLOAT: callq  __gnu_f2h_ieee
57; SOFTFLOAT: callq  __gnu_h2f_ieee
58; F16C: vcvtps2ph
59; F16C-NEXT: vcvtph2ps
60; F16C: ret
61
62define double @test4(i16* nocapture %src) {
63  %1 = load i16, i16* %src, align 2
64  %2 = tail call double @llvm.convert.from.fp16.f64(i16 %1)
65  ret double %2
66}
67; CHECK-LABEL: test4:
68; LIBCALL: callq  __gnu_h2f_ieee
69; LIBCALL: cvtss2sd
70; SOFTFLOAT: callq  __gnu_h2f_ieee
71; SOFTFLOAT: callq __extendsfdf2
72; F16C: vcvtph2ps
73; F16C: vcvtss2sd
74; F16C: ret
75
76
77define i16 @test5(double %src) {
78  %val = tail call i16 @llvm.convert.to.fp16.f64(double %src)
79  ret i16 %val
80}
81; CHECK-LABEL: test5:
82; LIBCALL: jmp  __truncdfhf2
83; SOFTFLOAT: callq  __truncdfhf2
84; F16C: jmp __truncdfhf2
85
86declare float @llvm.convert.from.fp16.f32(i16) nounwind readnone
87declare i16 @llvm.convert.to.fp16.f32(float) nounwind readnone
88declare double @llvm.convert.from.fp16.f64(i16) nounwind readnone
89declare i16 @llvm.convert.to.fp16.f64(double) nounwind readnone
90