1; RUN: llc -mtriple=armv7a--none-eabi < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-HARDFLOAT-EABI %s
2; RUN: llc -mtriple=armv7a--none-gnueabi < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-HARDFLOAT-GNU %s
3; RUN: llc -mtriple=armv7a--none-musleabi < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-HARDFLOAT-GNU %s
4; RUN: llc -mtriple=armv8-eabihf < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-ARMV8 %s
5; RUN: llc -mtriple=thumbv7m-eabi < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-SOFTFLOAT-EABI %s
6; RUN: llc -mtriple=thumbv7m-gnueabi < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-SOFTFLOAT-GNU %s
7; RUN: llc -mtriple=thumbv7m-musleabi < %s | FileCheck --check-prefix=CHECK --check-prefix=CHECK-SOFTFLOAT-GNU %s
8
9;; +fp16 is special: it has f32->f16 (unlike v7), but not f64->f16 (unlike v8).
10;; This exposes unsafe-fp-math optimization opportunities; test that.
11; RUN: llc -mattr=+vfp3,+fp16 < %s |\
12; RUN:   FileCheck --check-prefix=CHECK --check-prefix=CHECK-FP16 --check-prefix=CHECK-FP16-SAFE %s
13; RUN: llc -mattr=+vfp3,+fp16 < %s -enable-unsafe-fp-math |\
14; RUN:   FileCheck --check-prefix=CHECK --check-prefix=CHECK-FP16 --check-prefix=CHECK-FP16-UNSAFE %s
15
16target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-n32"
17target triple = "armv7---eabihf"
18
19@x = global i16 12902
20@y = global i16 0
21@z = common global i16 0
22
23define void @foo() nounwind {
24; CHECK-LABEL: foo:
25entry:
26  %0 = load i16, i16* @x, align 2
27  %1 = load i16, i16* @y, align 2
28  %2 = tail call float @llvm.convert.from.fp16.f32(i16 %0)
29; CHECK-HARDFLOAT-EABI: __aeabi_h2f
30; CHECK-HARDFLOAT-GNU: __gnu_h2f_ieee
31; CHECK-FP16: vcvtb.f32.f16
32; CHECK-ARMV8: vcvtb.f32.f16
33; CHECK-SOFTFLOAT-EABI: __aeabi_h2f
34; CHECK-SOFTFLOAT-GNU: __gnu_h2f_ieee
35  %3 = tail call float @llvm.convert.from.fp16.f32(i16 %1)
36; CHECK-HARDFLOAT-EABI: __aeabi_h2f
37; CHECK-HARDFLOAT-GNU: __gnu_h2f_ieee
38; CHECK-FP16: vcvtb.f32.f16
39; CHECK-ARMV8: vcvtb.f32.f16
40; CHECK-SOFTFLOAT-EABI: __aeabi_h2f
41; CHECK-SOFTFLOAT-GNU: __gnu_h2f_ieee
42  %4 = fadd float %2, %3
43  %5 = tail call i16 @llvm.convert.to.fp16.f32(float %4)
44; CHECK-HARDFLOAT-EABI: __aeabi_f2h
45; CHECK-HARDFLOAT-GNU: __gnu_f2h_ieee
46; CHECK-FP16: vcvtb.f16.f32
47; CHECK-ARMV8: vcvtb.f16.f32
48; CHECK-SOFTFLOAT-EABI: __aeabi_f2h
49; CHECK-SOFTFLOAT-GNU: __gnu_f2h_ieee
50  store i16 %5, i16* @x, align 2
51  ret void
52}
53
54define double @test_from_fp16(i16 %in) {
55; CHECK-LABEL: test_from_fp16:
56  %val = call double @llvm.convert.from.fp16.f64(i16 %in)
57; CHECK-HARDFLOAT-EABI: bl __aeabi_h2f
58; CHECK-HARDFLOAT-EABI: vmov [[TMP:s[0-9]+]], r0
59; CHECK-HARDFLOAT-EABI: vcvt.f64.f32 {{d[0-9]+}}, [[TMP]]
60
61; CHECK-HARDFLOAT-GNU: bl __gnu_h2f_ieee
62; CHECK-HARDFLOAT-GNU: vmov [[TMP:s[0-9]+]], r0
63; CHECK-HARDFLOAT-GNU: vcvt.f64.f32 {{d[0-9]+}}, [[TMP]]
64
65; CHECK-FP16: vmov [[TMP16:s[0-9]+]], r0
66; CHECK-FP16: vcvtb.f32.f16 [[TMP32:s[0-9]+]], [[TMP16]]
67; CHECK-FP16: vcvt.f64.f32 d0, [[TMP32]]
68
69; CHECK-ARMV8: vmov [[TMP:s[0-9]+]], r0
70; CHECK-ARMV8: vcvtb.f64.f16 d0, [[TMP]]
71
72; CHECK-SOFTFLOAT-EABI: bl __aeabi_h2f
73; CHECK-SOFTFLOAT-EABI: bl __aeabi_f2d
74
75; CHECK-SOFTFLOAT-GNU: bl __gnu_h2f_ieee
76; CHECK-SOFTFLOAT-GNU: bl __aeabi_f2d
77  ret double %val
78}
79
80define i16 @test_to_fp16(double %in) {
81; CHECK-LABEL: test_to_fp16:
82  %val = call i16 @llvm.convert.to.fp16.f64(double %in)
83; CHECK-HARDFLOAT-EABI: bl __aeabi_d2h
84
85; CHECK-HARDFLOAT-GNU: bl __aeabi_d2h
86
87; CHECK-FP16-SAFE: bl __aeabi_d2h
88
89; CHECK-FP16-UNSAFE:      vcvt.f32.f64 s0, d0
90; CHECK-FP16-UNSAFE-NEXT: vcvtb.f16.f32 s0, s0
91
92; CHECK-ARMV8: vcvtb.f16.f64 [[TMP:s[0-9]+]], d0
93; CHECK-ARMV8: vmov r0, [[TMP]]
94
95; CHECK-SOFTFLOAT-EABI: bl __aeabi_d2h
96
97; CHECK-SOFTFLOAT-GNU: bl __aeabi_d2h
98  ret i16 %val
99}
100
101declare float @llvm.convert.from.fp16.f32(i16) nounwind readnone
102declare double @llvm.convert.from.fp16.f64(i16) nounwind readnone
103
104declare i16 @llvm.convert.to.fp16.f32(float) nounwind readnone
105declare i16 @llvm.convert.to.fp16.f64(double) nounwind readnone
106