1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=x86_64-- -mcpu=x86-64 | FileCheck %s -check-prefix=X64
3; RUN: llc < %s -mtriple=i686-unknown   | FileCheck %s -check-prefix=X86
4
5declare float @llvm.sqrt.f32(float %x);
6
7define float @fast_recip_sqrt(float %x) {
8; X64-LABEL: fast_recip_sqrt:
9; X64:       # %bb.0:
10; X64-NEXT:    rsqrtss %xmm0, %xmm1
11; X64-NEXT:    mulss %xmm1, %xmm0
12; X64-NEXT:    mulss %xmm1, %xmm0
13; X64-NEXT:    addss {{.*}}(%rip), %xmm0
14; X64-NEXT:    mulss {{.*}}(%rip), %xmm1
15; X64-NEXT:    mulss %xmm1, %xmm0
16; X64-NEXT:    retq
17;
18; X86-LABEL: fast_recip_sqrt:
19; X86:       # %bb.0:
20; X86-NEXT:    flds {{[0-9]+}}(%esp)
21; X86-NEXT:    fsqrt
22; X86-NEXT:    fld1
23; X86-NEXT:    fdivp %st, %st(1)
24; X86-NEXT:    retl
25  %y = call fast float @llvm.sqrt.f32(float %x)
26  %z = fdiv fast float 1.0,  %y
27  ret float %z
28}
29
30declare float @llvm.fmuladd.f32(float %a, float %b, float %c);
31
32define float @fast_fmuladd_opts(float %a , float %b , float %c) {
33; X64-LABEL: fast_fmuladd_opts:
34; X64:       # %bb.0:
35; X64-NEXT:    mulss {{.*}}(%rip), %xmm0
36; X64-NEXT:    retq
37;
38; X86-LABEL: fast_fmuladd_opts:
39; X86:       # %bb.0:
40; X86-NEXT:    flds {{[0-9]+}}(%esp)
41; X86-NEXT:    fmuls {{\.LCPI.*}}
42; X86-NEXT:    retl
43  %res = call fast float @llvm.fmuladd.f32(float %a, float 2.0, float %a)
44  ret float %res
45}
46
47; The multiply is strict.
48
49@mul1 = common global double 0.000000e+00, align 4
50
51define double @not_so_fast_mul_add(double %x) {
52; X64-LABEL: not_so_fast_mul_add:
53; X64:       # %bb.0:
54; X64-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
55; X64-NEXT:    mulsd %xmm0, %xmm1
56; X64-NEXT:    mulsd {{.*}}(%rip), %xmm0
57; X64-NEXT:    movsd %xmm1, {{.*}}(%rip)
58; X64-NEXT:    retq
59;
60; X86-LABEL: not_so_fast_mul_add:
61; X86:       # %bb.0:
62; X86-NEXT:    fldl {{[0-9]+}}(%esp)
63; X86-NEXT:    fld %st(0)
64; X86-NEXT:    fmull {{\.LCPI.*}}
65; X86-NEXT:    fxch %st(1)
66; X86-NEXT:    fmull {{\.LCPI.*}}
67; X86-NEXT:    fxch %st(1)
68; X86-NEXT:    fstpl mul1
69; X86-NEXT:    retl
70  %m = fmul double %x, 4.2
71  %a = fadd fast double %m, %x
72  store double %m, double* @mul1, align 4
73  ret double %a
74}
75
76; The sqrt is strict.
77
78@sqrt1 = common global float 0.000000e+00, align 4
79
80define float @not_so_fast_recip_sqrt(float %x) {
81; X64-LABEL: not_so_fast_recip_sqrt:
82; X64:       # %bb.0:
83; X64-NEXT:    rsqrtss %xmm0, %xmm1
84; X64-NEXT:    sqrtss %xmm0, %xmm2
85; X64-NEXT:    mulss %xmm1, %xmm0
86; X64-NEXT:    mulss %xmm1, %xmm0
87; X64-NEXT:    addss {{.*}}(%rip), %xmm0
88; X64-NEXT:    mulss {{.*}}(%rip), %xmm1
89; X64-NEXT:    mulss %xmm1, %xmm0
90; X64-NEXT:    movss %xmm2, {{.*}}(%rip)
91; X64-NEXT:    retq
92;
93; X86-LABEL: not_so_fast_recip_sqrt:
94; X86:       # %bb.0:
95; X86-NEXT:    flds {{[0-9]+}}(%esp)
96; X86-NEXT:    fsqrt
97; X86-NEXT:    fld1
98; X86-NEXT:    fdiv %st(1), %st
99; X86-NEXT:    fxch %st(1)
100; X86-NEXT:    fstps sqrt1
101; X86-NEXT:    retl
102  %y = call float @llvm.sqrt.f32(float %x)
103  %z = fdiv fast float 1.0, %y
104  store float %y, float* @sqrt1, align 4
105  %ret = fadd float %z , 14.5
106  ret float %z
107}
108
109define float @div_arcp_by_const(half %x) {
110; X64-LABEL: div_arcp_by_const:
111; X64:       # %bb.0:
112; X64-NEXT:    pushq %rax
113; X64-NEXT:    .cfi_def_cfa_offset 16
114; X64-NEXT:    movzwl %di, %edi
115; X64-NEXT:    callq __gnu_h2f_ieee
116; X64-NEXT:    mulss {{.*}}(%rip), %xmm0
117; X64-NEXT:    callq __gnu_f2h_ieee
118; X64-NEXT:    movzwl %ax, %edi
119; X64-NEXT:    popq %rax
120; X64-NEXT:    .cfi_def_cfa_offset 8
121; X64-NEXT:    jmp __gnu_h2f_ieee # TAILCALL
122;
123; X86-LABEL: div_arcp_by_const:
124; X86:       # %bb.0:
125; X86-NEXT:    pushl %eax
126; X86-NEXT:    .cfi_def_cfa_offset 8
127; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
128; X86-NEXT:    movl %eax, (%esp)
129; X86-NEXT:    calll __gnu_h2f_ieee
130; X86-NEXT:    fmuls {{\.LCPI.*}}
131; X86-NEXT:    fstps (%esp)
132; X86-NEXT:    calll __gnu_f2h_ieee
133; X86-NEXT:    movzwl %ax, %eax
134; X86-NEXT:    movl %eax, (%esp)
135; X86-NEXT:    calll __gnu_h2f_ieee
136; X86-NEXT:    popl %eax
137; X86-NEXT:    .cfi_def_cfa_offset 4
138; X86-NEXT:    retl
139  %rcp = fdiv arcp half %x, 10.0
140  %z = fpext half %rcp to float
141  ret float %z
142}
143