1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -mtriple=x86_64-linux-gnu < %s  | FileCheck %s
3
4; Basic test coverage for FREM
5
6define void @frem_f16(half %a0, half %a1, half *%p3) nounwind {
7; CHECK-LABEL: frem_f16:
8; CHECK:       # %bb.0:
9; CHECK-NEXT:    pushq %rbp
10; CHECK-NEXT:    pushq %rbx
11; CHECK-NEXT:    pushq %rax
12; CHECK-NEXT:    movq %rdx, %rbx
13; CHECK-NEXT:    movl %edi, %ebp
14; CHECK-NEXT:    movzwl %si, %edi
15; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
16; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
17; CHECK-NEXT:    movzwl %bp, %edi
18; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
19; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
20; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
21; CHECK-NEXT:    callq fmodf@PLT
22; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
23; CHECK-NEXT:    movw %ax, (%rbx)
24; CHECK-NEXT:    addq $8, %rsp
25; CHECK-NEXT:    popq %rbx
26; CHECK-NEXT:    popq %rbp
27; CHECK-NEXT:    retq
28  %frem = frem half %a0, %a1
29  store half %frem, half *%p3
30  ret void
31}
32
33define void @frem_f32(float %a0, float %a1, float *%p3) nounwind {
34; CHECK-LABEL: frem_f32:
35; CHECK:       # %bb.0:
36; CHECK-NEXT:    pushq %rbx
37; CHECK-NEXT:    movq %rdi, %rbx
38; CHECK-NEXT:    callq fmodf@PLT
39; CHECK-NEXT:    movss %xmm0, (%rbx)
40; CHECK-NEXT:    popq %rbx
41; CHECK-NEXT:    retq
42  %frem = frem float %a0, %a1
43  store float %frem, float *%p3
44  ret void
45}
46
47define void @frem_f64(double %a0, double %a1, double *%p3) nounwind {
48; CHECK-LABEL: frem_f64:
49; CHECK:       # %bb.0:
50; CHECK-NEXT:    pushq %rbx
51; CHECK-NEXT:    movq %rdi, %rbx
52; CHECK-NEXT:    callq fmod@PLT
53; CHECK-NEXT:    movsd %xmm0, (%rbx)
54; CHECK-NEXT:    popq %rbx
55; CHECK-NEXT:    retq
56  %frem = frem double %a0, %a1
57  store double %frem, double *%p3
58  ret void
59}
60
61define void @frem_f80(x86_fp80 %a0, x86_fp80 %a1, x86_fp80 *%p3) nounwind {
62; CHECK-LABEL: frem_f80:
63; CHECK:       # %bb.0:
64; CHECK-NEXT:    pushq %rbx
65; CHECK-NEXT:    subq $32, %rsp
66; CHECK-NEXT:    movq %rdi, %rbx
67; CHECK-NEXT:    fldt {{[0-9]+}}(%rsp)
68; CHECK-NEXT:    fldt {{[0-9]+}}(%rsp)
69; CHECK-NEXT:    fstpt {{[0-9]+}}(%rsp)
70; CHECK-NEXT:    fstpt (%rsp)
71; CHECK-NEXT:    callq fmodl@PLT
72; CHECK-NEXT:    fstpt (%rbx)
73; CHECK-NEXT:    addq $32, %rsp
74; CHECK-NEXT:    popq %rbx
75; CHECK-NEXT:    retq
76  %frem = frem x86_fp80 %a0, %a1
77  store x86_fp80 %frem, x86_fp80 *%p3
78  ret void
79}
80
81define void @frem_f128(fp128 %a0, fp128 %a1, fp128 *%p3) nounwind {
82; CHECK-LABEL: frem_f128:
83; CHECK:       # %bb.0:
84; CHECK-NEXT:    pushq %rbx
85; CHECK-NEXT:    movq %rdi, %rbx
86; CHECK-NEXT:    callq fmodl@PLT
87; CHECK-NEXT:    movaps %xmm0, (%rbx)
88; CHECK-NEXT:    popq %rbx
89; CHECK-NEXT:    retq
90  %frem = frem fp128 %a0, %a1
91  store fp128 %frem, fp128 *%p3
92  ret void
93}
94
95define void @frem_v16f32(<16 x float> %a0, <16 x float> %a1, <16 x float> *%p3) nounwind {
96; CHECK-LABEL: frem_v16f32:
97; CHECK:       # %bb.0:
98; CHECK-NEXT:    pushq %rbx
99; CHECK-NEXT:    subq $160, %rsp
100; CHECK-NEXT:    movq %rdi, %rbx
101; CHECK-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
102; CHECK-NEXT:    movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
103; CHECK-NEXT:    movaps %xmm5, (%rsp) # 16-byte Spill
104; CHECK-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
105; CHECK-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
106; CHECK-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
107; CHECK-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
108; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
109; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
110; CHECK-NEXT:    movaps %xmm4, %xmm1
111; CHECK-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,3],xmm4[3,3]
112; CHECK-NEXT:    callq fmodf@PLT
113; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
114; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
115; CHECK-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
116; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
117; CHECK-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
118; CHECK-NEXT:    callq fmodf@PLT
119; CHECK-NEXT:    unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
120; CHECK-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
121; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
122; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
123; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
124; CHECK-NEXT:    callq fmodf@PLT
125; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
126; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
127; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
128; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
129; CHECK-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,1,1,1]
130; CHECK-NEXT:    callq fmodf@PLT
131; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
132; CHECK-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
133; CHECK-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
134; CHECK-NEXT:    # xmm1 = xmm1[0],mem[0]
135; CHECK-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
136; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
137; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
138; CHECK-NEXT:    movaps (%rsp), %xmm1 # 16-byte Reload
139; CHECK-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,3,3,3]
140; CHECK-NEXT:    callq fmodf@PLT
141; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
142; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
143; CHECK-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
144; CHECK-NEXT:    movaps (%rsp), %xmm1 # 16-byte Reload
145; CHECK-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
146; CHECK-NEXT:    callq fmodf@PLT
147; CHECK-NEXT:    unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
148; CHECK-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
149; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
150; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
151; CHECK-NEXT:    movaps (%rsp), %xmm1 # 16-byte Reload
152; CHECK-NEXT:    callq fmodf@PLT
153; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
154; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
155; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
156; CHECK-NEXT:    movaps (%rsp), %xmm1 # 16-byte Reload
157; CHECK-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,1,1,1]
158; CHECK-NEXT:    callq fmodf@PLT
159; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
160; CHECK-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
161; CHECK-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
162; CHECK-NEXT:    # xmm1 = xmm1[0],mem[0]
163; CHECK-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
164; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
165; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
166; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
167; CHECK-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,3,3,3]
168; CHECK-NEXT:    callq fmodf@PLT
169; CHECK-NEXT:    movaps %xmm0, (%rsp) # 16-byte Spill
170; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
171; CHECK-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
172; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
173; CHECK-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
174; CHECK-NEXT:    callq fmodf@PLT
175; CHECK-NEXT:    unpcklps (%rsp), %xmm0 # 16-byte Folded Reload
176; CHECK-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
177; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
178; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
179; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
180; CHECK-NEXT:    callq fmodf@PLT
181; CHECK-NEXT:    movaps %xmm0, (%rsp) # 16-byte Spill
182; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
183; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
184; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
185; CHECK-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,1,1,1]
186; CHECK-NEXT:    callq fmodf@PLT
187; CHECK-NEXT:    movaps (%rsp), %xmm1 # 16-byte Reload
188; CHECK-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
189; CHECK-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
190; CHECK-NEXT:    # xmm1 = xmm1[0],mem[0]
191; CHECK-NEXT:    movaps %xmm1, (%rsp) # 16-byte Spill
192; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
193; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
194; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
195; CHECK-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,3,3,3]
196; CHECK-NEXT:    callq fmodf@PLT
197; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
198; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
199; CHECK-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
200; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
201; CHECK-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
202; CHECK-NEXT:    callq fmodf@PLT
203; CHECK-NEXT:    unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
204; CHECK-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
205; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
206; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
207; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
208; CHECK-NEXT:    callq fmodf@PLT
209; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
210; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
211; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
212; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
213; CHECK-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,1,1,1]
214; CHECK-NEXT:    callq fmodf@PLT
215; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
216; CHECK-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
217; CHECK-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
218; CHECK-NEXT:    # xmm1 = xmm1[0],mem[0]
219; CHECK-NEXT:    movaps %xmm1, 48(%rbx)
220; CHECK-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload
221; CHECK-NEXT:    movaps %xmm0, 32(%rbx)
222; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
223; CHECK-NEXT:    movaps %xmm0, 16(%rbx)
224; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
225; CHECK-NEXT:    movaps %xmm0, (%rbx)
226; CHECK-NEXT:    addq $160, %rsp
227; CHECK-NEXT:    popq %rbx
228; CHECK-NEXT:    retq
229  %frem = frem <16 x float> %a0, %a1
230  store <16 x float> %frem, <16 x float> *%p3
231  ret void
232}
233
234define void @frem_v8f32(<8 x float> %a0, <8 x float> %a1, <8 x float> *%p3) nounwind {
235; CHECK-LABEL: frem_v8f32:
236; CHECK:       # %bb.0:
237; CHECK-NEXT:    pushq %rbx
238; CHECK-NEXT:    subq $96, %rsp
239; CHECK-NEXT:    movq %rdi, %rbx
240; CHECK-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
241; CHECK-NEXT:    movaps %xmm2, (%rsp) # 16-byte Spill
242; CHECK-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
243; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
244; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
245; CHECK-NEXT:    movaps %xmm2, %xmm1
246; CHECK-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,3],xmm2[3,3]
247; CHECK-NEXT:    callq fmodf@PLT
248; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
249; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
250; CHECK-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
251; CHECK-NEXT:    movaps (%rsp), %xmm1 # 16-byte Reload
252; CHECK-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
253; CHECK-NEXT:    callq fmodf@PLT
254; CHECK-NEXT:    unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
255; CHECK-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
256; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
257; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
258; CHECK-NEXT:    movaps (%rsp), %xmm1 # 16-byte Reload
259; CHECK-NEXT:    callq fmodf@PLT
260; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
261; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
262; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
263; CHECK-NEXT:    movaps (%rsp), %xmm1 # 16-byte Reload
264; CHECK-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,1,1,1]
265; CHECK-NEXT:    callq fmodf@PLT
266; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
267; CHECK-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
268; CHECK-NEXT:    unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
269; CHECK-NEXT:    # xmm1 = xmm1[0],mem[0]
270; CHECK-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
271; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
272; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
273; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
274; CHECK-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,3,3,3]
275; CHECK-NEXT:    callq fmodf@PLT
276; CHECK-NEXT:    movaps %xmm0, (%rsp) # 16-byte Spill
277; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
278; CHECK-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
279; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
280; CHECK-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
281; CHECK-NEXT:    callq fmodf@PLT
282; CHECK-NEXT:    unpcklps (%rsp), %xmm0 # 16-byte Folded Reload
283; CHECK-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
284; CHECK-NEXT:    movaps %xmm0, (%rsp) # 16-byte Spill
285; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
286; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
287; CHECK-NEXT:    callq fmodf@PLT
288; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
289; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
290; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
291; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
292; CHECK-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,1,1,1]
293; CHECK-NEXT:    callq fmodf@PLT
294; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
295; CHECK-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
296; CHECK-NEXT:    unpcklpd (%rsp), %xmm1 # 16-byte Folded Reload
297; CHECK-NEXT:    # xmm1 = xmm1[0],mem[0]
298; CHECK-NEXT:    movaps %xmm1, 16(%rbx)
299; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
300; CHECK-NEXT:    movaps %xmm0, (%rbx)
301; CHECK-NEXT:    addq $96, %rsp
302; CHECK-NEXT:    popq %rbx
303; CHECK-NEXT:    retq
304  %frem = frem <8 x float> %a0, %a1
305  store <8 x float> %frem, <8 x float> *%p3
306  ret void
307}
308
309define void @frem_v4f32(<4 x float> %a0, <4 x float> %a1, <4 x float> *%p3) nounwind {
310; CHECK-LABEL: frem_v4f32:
311; CHECK:       # %bb.0:
312; CHECK-NEXT:    pushq %rbx
313; CHECK-NEXT:    subq $64, %rsp
314; CHECK-NEXT:    movq %rdi, %rbx
315; CHECK-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
316; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
317; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
318; CHECK-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,3,3,3]
319; CHECK-NEXT:    callq fmodf@PLT
320; CHECK-NEXT:    movaps %xmm0, (%rsp) # 16-byte Spill
321; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
322; CHECK-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
323; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
324; CHECK-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
325; CHECK-NEXT:    callq fmodf@PLT
326; CHECK-NEXT:    unpcklps (%rsp), %xmm0 # 16-byte Folded Reload
327; CHECK-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
328; CHECK-NEXT:    movaps %xmm0, (%rsp) # 16-byte Spill
329; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
330; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
331; CHECK-NEXT:    callq fmodf@PLT
332; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
333; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
334; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
335; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
336; CHECK-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,1,1,1]
337; CHECK-NEXT:    callq fmodf@PLT
338; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
339; CHECK-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
340; CHECK-NEXT:    unpcklpd (%rsp), %xmm1 # 16-byte Folded Reload
341; CHECK-NEXT:    # xmm1 = xmm1[0],mem[0]
342; CHECK-NEXT:    movaps %xmm1, (%rbx)
343; CHECK-NEXT:    addq $64, %rsp
344; CHECK-NEXT:    popq %rbx
345; CHECK-NEXT:    retq
346  %frem = frem <4 x float> %a0, %a1
347  store <4 x float> %frem, <4 x float> *%p3
348  ret void
349}
350
351define void @frem_v8f64(<8 x double> %a0, <8 x double> %a1, <8 x double> *%p3) nounwind {
352; CHECK-LABEL: frem_v8f64:
353; CHECK:       # %bb.0:
354; CHECK-NEXT:    pushq %rbx
355; CHECK-NEXT:    subq $144, %rsp
356; CHECK-NEXT:    movq %rdi, %rbx
357; CHECK-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
358; CHECK-NEXT:    movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
359; CHECK-NEXT:    movaps %xmm5, (%rsp) # 16-byte Spill
360; CHECK-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
361; CHECK-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
362; CHECK-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
363; CHECK-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
364; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
365; CHECK-NEXT:    movaps %xmm4, %xmm1
366; CHECK-NEXT:    callq fmod@PLT
367; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
368; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
369; CHECK-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
370; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
371; CHECK-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
372; CHECK-NEXT:    callq fmod@PLT
373; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
374; CHECK-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
375; CHECK-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
376; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
377; CHECK-NEXT:    movaps (%rsp), %xmm1 # 16-byte Reload
378; CHECK-NEXT:    callq fmod@PLT
379; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
380; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
381; CHECK-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
382; CHECK-NEXT:    movaps (%rsp), %xmm1 # 16-byte Reload
383; CHECK-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
384; CHECK-NEXT:    callq fmod@PLT
385; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
386; CHECK-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
387; CHECK-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
388; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
389; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
390; CHECK-NEXT:    callq fmod@PLT
391; CHECK-NEXT:    movaps %xmm0, (%rsp) # 16-byte Spill
392; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
393; CHECK-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
394; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
395; CHECK-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
396; CHECK-NEXT:    callq fmod@PLT
397; CHECK-NEXT:    movaps (%rsp), %xmm1 # 16-byte Reload
398; CHECK-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
399; CHECK-NEXT:    movaps %xmm1, (%rsp) # 16-byte Spill
400; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
401; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
402; CHECK-NEXT:    callq fmod@PLT
403; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
404; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
405; CHECK-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
406; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
407; CHECK-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
408; CHECK-NEXT:    callq fmod@PLT
409; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
410; CHECK-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
411; CHECK-NEXT:    movaps %xmm1, 48(%rbx)
412; CHECK-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload
413; CHECK-NEXT:    movaps %xmm0, 32(%rbx)
414; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
415; CHECK-NEXT:    movaps %xmm0, 16(%rbx)
416; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
417; CHECK-NEXT:    movaps %xmm0, (%rbx)
418; CHECK-NEXT:    addq $144, %rsp
419; CHECK-NEXT:    popq %rbx
420; CHECK-NEXT:    retq
421  %frem = frem <8 x double> %a0, %a1
422  store <8 x double> %frem, <8 x double> *%p3
423  ret void
424}
425
426define void @frem_v4f64(<4 x double> %a0, <4 x double> %a1, <4 x double> *%p3) nounwind {
427; CHECK-LABEL: frem_v4f64:
428; CHECK:       # %bb.0:
429; CHECK-NEXT:    pushq %rbx
430; CHECK-NEXT:    subq $80, %rsp
431; CHECK-NEXT:    movq %rdi, %rbx
432; CHECK-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
433; CHECK-NEXT:    movaps %xmm2, (%rsp) # 16-byte Spill
434; CHECK-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
435; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
436; CHECK-NEXT:    movaps %xmm2, %xmm1
437; CHECK-NEXT:    callq fmod@PLT
438; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
439; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
440; CHECK-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
441; CHECK-NEXT:    movaps (%rsp), %xmm1 # 16-byte Reload
442; CHECK-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
443; CHECK-NEXT:    callq fmod@PLT
444; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
445; CHECK-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
446; CHECK-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
447; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
448; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
449; CHECK-NEXT:    callq fmod@PLT
450; CHECK-NEXT:    movaps %xmm0, (%rsp) # 16-byte Spill
451; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
452; CHECK-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
453; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
454; CHECK-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
455; CHECK-NEXT:    callq fmod@PLT
456; CHECK-NEXT:    movaps (%rsp), %xmm1 # 16-byte Reload
457; CHECK-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
458; CHECK-NEXT:    movaps %xmm1, 16(%rbx)
459; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
460; CHECK-NEXT:    movaps %xmm0, (%rbx)
461; CHECK-NEXT:    addq $80, %rsp
462; CHECK-NEXT:    popq %rbx
463; CHECK-NEXT:    retq
464  %frem = frem <4 x double> %a0, %a1
465  store <4 x double> %frem, <4 x double> *%p3
466  ret void
467}
468
469define void @frem_v2f64(<2 x double> %a0, <2 x double> %a1, <2 x double> *%p3) nounwind {
470; CHECK-LABEL: frem_v2f64:
471; CHECK:       # %bb.0:
472; CHECK-NEXT:    pushq %rbx
473; CHECK-NEXT:    subq $48, %rsp
474; CHECK-NEXT:    movq %rdi, %rbx
475; CHECK-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
476; CHECK-NEXT:    movaps %xmm0, (%rsp) # 16-byte Spill
477; CHECK-NEXT:    callq fmod@PLT
478; CHECK-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
479; CHECK-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload
480; CHECK-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
481; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
482; CHECK-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
483; CHECK-NEXT:    callq fmod@PLT
484; CHECK-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
485; CHECK-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
486; CHECK-NEXT:    movaps %xmm1, (%rbx)
487; CHECK-NEXT:    addq $48, %rsp
488; CHECK-NEXT:    popq %rbx
489; CHECK-NEXT:    retq
490  %frem = frem <2 x double> %a0, %a1
491  store <2 x double> %frem, <2 x double> *%p3
492  ret void
493}
494
495define void @frem_v32f16(<32 x half> %a0, <32 x half> %a1, <32 x half> *%p3) nounwind {
496; CHECK-LABEL: frem_v32f16:
497; CHECK:       # %bb.0:
498; CHECK-NEXT:    pushq %rbp
499; CHECK-NEXT:    pushq %r15
500; CHECK-NEXT:    pushq %r14
501; CHECK-NEXT:    pushq %r13
502; CHECK-NEXT:    pushq %r12
503; CHECK-NEXT:    pushq %rbx
504; CHECK-NEXT:    subq $248, %rsp
505; CHECK-NEXT:    movl %r9d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
506; CHECK-NEXT:    movl %r8d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
507; CHECK-NEXT:    movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
508; CHECK-NEXT:    movl %edx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
509; CHECK-NEXT:    movl %esi, %r14d
510; CHECK-NEXT:    movl %edi, %ebx
511; CHECK-NEXT:    movq {{[0-9]+}}(%rsp), %r15
512; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
513; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
514; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
515; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
516; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
517; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
518; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
519; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
520; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
521; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
522; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
523; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
524; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
525; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
526; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
527; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
528; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
529; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
530; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
531; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
532; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
533; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
534; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
535; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
536; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
537; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
538; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
539; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
540; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
541; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
542; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
543; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
544; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
545; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
546; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
547; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
548; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
549; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
550; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
551; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
552; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
553; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
554; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
555; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
556; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
557; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
558; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
559; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
560; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
561; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
562; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
563; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
564; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
565; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
566; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
567; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
568; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
569; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
570; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
571; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
572; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
573; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
574; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
575; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
576; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
577; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
578; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
579; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
580; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
581; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
582; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
583; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
584; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
585; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
586; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
587; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
588; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
589; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
590; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
591; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
592; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
593; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
594; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
595; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
596; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
597; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
598; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
599; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
600; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
601; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
602; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
603; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
604; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
605; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
606; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
607; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
608; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
609; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
610; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
611; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
612; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
613; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
614; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
615; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
616; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
617; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
618; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
619; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
620; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %ebp
621; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %r13d
622; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %r12d
623; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %edi
624; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
625; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
626; CHECK-NEXT:    movzwl %bx, %edi
627; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
628; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
629; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
630; CHECK-NEXT:    callq fmodf@PLT
631; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
632; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
633; CHECK-NEXT:    movl %r12d, %edi
634; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
635; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
636; CHECK-NEXT:    movzwl %r14w, %edi
637; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
638; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
639; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
640; CHECK-NEXT:    callq fmodf@PLT
641; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
642; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
643; CHECK-NEXT:    movl %r13d, %edi
644; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
645; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
646; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 2-byte Folded Reload
647; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
648; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
649; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
650; CHECK-NEXT:    callq fmodf@PLT
651; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
652; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
653; CHECK-NEXT:    movl %ebp, %edi
654; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
655; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
656; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 2-byte Folded Reload
657; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
658; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
659; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
660; CHECK-NEXT:    callq fmodf@PLT
661; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
662; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
663; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
664; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
665; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
666; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 2-byte Folded Reload
667; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
668; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
669; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
670; CHECK-NEXT:    callq fmodf@PLT
671; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
672; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
673; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
674; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
675; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
676; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 2-byte Folded Reload
677; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
678; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
679; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
680; CHECK-NEXT:    callq fmodf@PLT
681; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
682; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
683; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
684; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
685; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
686; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
687; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
688; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
689; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
690; CHECK-NEXT:    callq fmodf@PLT
691; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
692; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
693; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
694; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
695; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
696; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
697; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
698; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
699; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
700; CHECK-NEXT:    callq fmodf@PLT
701; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
702; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
703; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
704; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
705; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
706; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
707; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
708; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
709; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
710; CHECK-NEXT:    callq fmodf@PLT
711; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
712; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
713; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
714; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
715; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
716; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
717; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
718; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
719; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
720; CHECK-NEXT:    callq fmodf@PLT
721; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
722; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
723; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
724; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
725; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
726; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
727; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
728; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
729; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
730; CHECK-NEXT:    callq fmodf@PLT
731; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
732; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
733; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
734; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
735; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
736; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
737; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
738; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
739; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
740; CHECK-NEXT:    callq fmodf@PLT
741; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
742; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
743; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
744; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
745; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
746; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
747; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
748; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
749; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
750; CHECK-NEXT:    callq fmodf@PLT
751; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
752; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
753; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
754; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
755; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
756; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
757; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
758; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
759; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
760; CHECK-NEXT:    callq fmodf@PLT
761; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
762; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
763; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
764; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
765; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
766; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
767; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
768; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
769; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
770; CHECK-NEXT:    callq fmodf@PLT
771; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
772; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
773; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
774; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
775; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
776; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
777; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
778; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
779; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
780; CHECK-NEXT:    callq fmodf@PLT
781; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
782; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
783; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
784; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
785; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
786; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
787; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
788; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
789; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
790; CHECK-NEXT:    callq fmodf@PLT
791; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
792; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
793; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
794; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
795; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
796; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
797; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
798; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
799; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
800; CHECK-NEXT:    callq fmodf@PLT
801; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
802; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
803; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
804; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
805; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
806; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
807; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
808; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
809; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
810; CHECK-NEXT:    callq fmodf@PLT
811; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
812; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
813; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
814; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
815; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
816; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
817; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
818; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
819; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
820; CHECK-NEXT:    callq fmodf@PLT
821; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
822; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
823; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
824; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
825; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
826; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
827; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
828; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
829; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
830; CHECK-NEXT:    callq fmodf@PLT
831; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
832; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
833; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
834; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
835; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
836; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
837; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
838; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
839; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
840; CHECK-NEXT:    callq fmodf@PLT
841; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
842; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
843; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
844; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
845; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
846; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
847; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
848; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
849; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
850; CHECK-NEXT:    callq fmodf@PLT
851; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
852; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
853; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
854; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
855; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
856; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
857; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
858; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
859; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
860; CHECK-NEXT:    callq fmodf@PLT
861; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
862; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
863; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
864; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
865; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
866; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
867; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
868; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
869; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
870; CHECK-NEXT:    callq fmodf@PLT
871; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
872; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
873; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
874; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
875; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
876; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
877; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
878; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
879; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
880; CHECK-NEXT:    callq fmodf@PLT
881; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
882; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
883; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
884; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
885; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
886; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
887; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
888; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
889; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
890; CHECK-NEXT:    callq fmodf@PLT
891; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
892; CHECK-NEXT:    movl %eax, %ebp
893; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
894; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
895; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
896; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
897; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
898; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
899; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
900; CHECK-NEXT:    callq fmodf@PLT
901; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
902; CHECK-NEXT:    movl %eax, %r14d
903; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
904; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
905; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
906; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
907; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
908; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
909; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
910; CHECK-NEXT:    callq fmodf@PLT
911; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
912; CHECK-NEXT:    movl %eax, %r12d
913; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
914; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
915; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
916; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
917; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
918; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
919; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
920; CHECK-NEXT:    callq fmodf@PLT
921; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
922; CHECK-NEXT:    movl %eax, %r13d
923; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
924; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
925; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
926; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
927; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
928; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
929; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
930; CHECK-NEXT:    callq fmodf@PLT
931; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
932; CHECK-NEXT:    movl %eax, %ebx
933; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
934; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
935; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
936; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
937; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
938; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
939; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
940; CHECK-NEXT:    callq fmodf@PLT
941; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
942; CHECK-NEXT:    movw %ax, 62(%r15)
943; CHECK-NEXT:    movw %bx, 60(%r15)
944; CHECK-NEXT:    movw %r13w, 58(%r15)
945; CHECK-NEXT:    movw %r12w, 56(%r15)
946; CHECK-NEXT:    movw %r14w, 54(%r15)
947; CHECK-NEXT:    movw %bp, 52(%r15)
948; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
949; CHECK-NEXT:    movw %ax, 50(%r15)
950; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
951; CHECK-NEXT:    movw %ax, 48(%r15)
952; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
953; CHECK-NEXT:    movw %ax, 46(%r15)
954; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
955; CHECK-NEXT:    movw %ax, 44(%r15)
956; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
957; CHECK-NEXT:    movw %ax, 42(%r15)
958; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
959; CHECK-NEXT:    movw %ax, 40(%r15)
960; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
961; CHECK-NEXT:    movw %ax, 38(%r15)
962; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
963; CHECK-NEXT:    movw %ax, 36(%r15)
964; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
965; CHECK-NEXT:    movw %ax, 34(%r15)
966; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
967; CHECK-NEXT:    movw %ax, 32(%r15)
968; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
969; CHECK-NEXT:    movw %ax, 30(%r15)
970; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
971; CHECK-NEXT:    movw %ax, 28(%r15)
972; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
973; CHECK-NEXT:    movw %ax, 26(%r15)
974; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
975; CHECK-NEXT:    movw %ax, 24(%r15)
976; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
977; CHECK-NEXT:    movw %ax, 22(%r15)
978; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
979; CHECK-NEXT:    movw %ax, 20(%r15)
980; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
981; CHECK-NEXT:    movw %ax, 18(%r15)
982; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
983; CHECK-NEXT:    movw %ax, 16(%r15)
984; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
985; CHECK-NEXT:    movw %ax, 14(%r15)
986; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
987; CHECK-NEXT:    movw %ax, 12(%r15)
988; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
989; CHECK-NEXT:    movw %ax, 10(%r15)
990; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
991; CHECK-NEXT:    movw %ax, 8(%r15)
992; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
993; CHECK-NEXT:    movw %ax, 6(%r15)
994; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
995; CHECK-NEXT:    movw %ax, 4(%r15)
996; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
997; CHECK-NEXT:    movw %ax, 2(%r15)
998; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
999; CHECK-NEXT:    movw %ax, (%r15)
1000; CHECK-NEXT:    addq $248, %rsp
1001; CHECK-NEXT:    popq %rbx
1002; CHECK-NEXT:    popq %r12
1003; CHECK-NEXT:    popq %r13
1004; CHECK-NEXT:    popq %r14
1005; CHECK-NEXT:    popq %r15
1006; CHECK-NEXT:    popq %rbp
1007; CHECK-NEXT:    retq
1008  %frem = frem <32 x half> %a0, %a1
1009  store <32 x half> %frem, <32 x half> *%p3
1010  ret void
1011}
1012
1013define void @frem_v16f16(<16 x half> %a0, <16 x half> %a1, <16 x half> *%p3) nounwind {
1014; CHECK-LABEL: frem_v16f16:
1015; CHECK:       # %bb.0:
1016; CHECK-NEXT:    pushq %rbp
1017; CHECK-NEXT:    pushq %r15
1018; CHECK-NEXT:    pushq %r14
1019; CHECK-NEXT:    pushq %r13
1020; CHECK-NEXT:    pushq %r12
1021; CHECK-NEXT:    pushq %rbx
1022; CHECK-NEXT:    subq $120, %rsp
1023; CHECK-NEXT:    movl %r9d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1024; CHECK-NEXT:    movl %r8d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1025; CHECK-NEXT:    movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1026; CHECK-NEXT:    movl %edx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1027; CHECK-NEXT:    movl %esi, %r15d
1028; CHECK-NEXT:    movl %edi, %r14d
1029; CHECK-NEXT:    movq {{[0-9]+}}(%rsp), %rbp
1030; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1031; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1032; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1033; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1034; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1035; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1036; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1037; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1038; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1039; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1040; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1041; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1042; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1043; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1044; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1045; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1046; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1047; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1048; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1049; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1050; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1051; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1052; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1053; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1054; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1055; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1056; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1057; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1058; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1059; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1060; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1061; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1062; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1063; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1064; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1065; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1066; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1067; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1068; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1069; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1070; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1071; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1072; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1073; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1074; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %r13d
1075; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %r12d
1076; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %ebx
1077; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %edi
1078; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1079; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1080; CHECK-NEXT:    movzwl %r14w, %edi
1081; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1082; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1083; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1084; CHECK-NEXT:    callq fmodf@PLT
1085; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1086; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
1087; CHECK-NEXT:    movl %ebx, %edi
1088; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1089; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1090; CHECK-NEXT:    movzwl %r15w, %edi
1091; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1092; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1093; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1094; CHECK-NEXT:    callq fmodf@PLT
1095; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1096; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
1097; CHECK-NEXT:    movl %r12d, %edi
1098; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1099; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1100; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 2-byte Folded Reload
1101; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1102; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1103; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1104; CHECK-NEXT:    callq fmodf@PLT
1105; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1106; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
1107; CHECK-NEXT:    movl %r13d, %edi
1108; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1109; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1110; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 2-byte Folded Reload
1111; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1112; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1113; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1114; CHECK-NEXT:    callq fmodf@PLT
1115; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1116; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
1117; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1118; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1119; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1120; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 2-byte Folded Reload
1121; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1122; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1123; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1124; CHECK-NEXT:    callq fmodf@PLT
1125; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1126; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
1127; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1128; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1129; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1130; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 2-byte Folded Reload
1131; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1132; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1133; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1134; CHECK-NEXT:    callq fmodf@PLT
1135; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1136; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
1137; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1138; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1139; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1140; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1141; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1142; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1143; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1144; CHECK-NEXT:    callq fmodf@PLT
1145; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1146; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
1147; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1148; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1149; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1150; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1151; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1152; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1153; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1154; CHECK-NEXT:    callq fmodf@PLT
1155; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1156; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
1157; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1158; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1159; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1160; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1161; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1162; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1163; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1164; CHECK-NEXT:    callq fmodf@PLT
1165; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1166; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
1167; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1168; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1169; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1170; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1171; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1172; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1173; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1174; CHECK-NEXT:    callq fmodf@PLT
1175; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1176; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
1177; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1178; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1179; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1180; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1181; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1182; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1183; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1184; CHECK-NEXT:    callq fmodf@PLT
1185; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1186; CHECK-NEXT:    movl %eax, %r13d
1187; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1188; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1189; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1190; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1191; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1192; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1193; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1194; CHECK-NEXT:    callq fmodf@PLT
1195; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1196; CHECK-NEXT:    movl %eax, %r12d
1197; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1198; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1199; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1200; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1201; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1202; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1203; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1204; CHECK-NEXT:    callq fmodf@PLT
1205; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1206; CHECK-NEXT:    movl %eax, %ebx
1207; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1208; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1209; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1210; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1211; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1212; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1213; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1214; CHECK-NEXT:    callq fmodf@PLT
1215; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1216; CHECK-NEXT:    movl %eax, %r14d
1217; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1218; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1219; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1220; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1221; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1222; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1223; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1224; CHECK-NEXT:    callq fmodf@PLT
1225; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1226; CHECK-NEXT:    movl %eax, %r15d
1227; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1228; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1229; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1230; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1231; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1232; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1233; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1234; CHECK-NEXT:    callq fmodf@PLT
1235; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1236; CHECK-NEXT:    movw %ax, 30(%rbp)
1237; CHECK-NEXT:    movw %r15w, 28(%rbp)
1238; CHECK-NEXT:    movw %r14w, 26(%rbp)
1239; CHECK-NEXT:    movw %bx, 24(%rbp)
1240; CHECK-NEXT:    movw %r12w, 22(%rbp)
1241; CHECK-NEXT:    movw %r13w, 20(%rbp)
1242; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
1243; CHECK-NEXT:    movw %ax, 18(%rbp)
1244; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
1245; CHECK-NEXT:    movw %ax, 16(%rbp)
1246; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
1247; CHECK-NEXT:    movw %ax, 14(%rbp)
1248; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
1249; CHECK-NEXT:    movw %ax, 12(%rbp)
1250; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
1251; CHECK-NEXT:    movw %ax, 10(%rbp)
1252; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
1253; CHECK-NEXT:    movw %ax, 8(%rbp)
1254; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
1255; CHECK-NEXT:    movw %ax, 6(%rbp)
1256; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
1257; CHECK-NEXT:    movw %ax, 4(%rbp)
1258; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
1259; CHECK-NEXT:    movw %ax, 2(%rbp)
1260; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
1261; CHECK-NEXT:    movw %ax, (%rbp)
1262; CHECK-NEXT:    addq $120, %rsp
1263; CHECK-NEXT:    popq %rbx
1264; CHECK-NEXT:    popq %r12
1265; CHECK-NEXT:    popq %r13
1266; CHECK-NEXT:    popq %r14
1267; CHECK-NEXT:    popq %r15
1268; CHECK-NEXT:    popq %rbp
1269; CHECK-NEXT:    retq
1270  %frem = frem <16 x half> %a0, %a1
1271  store <16 x half> %frem, <16 x half> *%p3
1272  ret void
1273}
1274
1275define void @frem_v8f16(<8 x half> %a0, <8 x half> %a1, <8 x half> *%p3) nounwind {
1276; CHECK-LABEL: frem_v8f16:
1277; CHECK:       # %bb.0:
1278; CHECK-NEXT:    pushq %rbp
1279; CHECK-NEXT:    pushq %r15
1280; CHECK-NEXT:    pushq %r14
1281; CHECK-NEXT:    pushq %r13
1282; CHECK-NEXT:    pushq %r12
1283; CHECK-NEXT:    pushq %rbx
1284; CHECK-NEXT:    subq $56, %rsp
1285; CHECK-NEXT:    movl %r9d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1286; CHECK-NEXT:    movl %r8d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1287; CHECK-NEXT:    movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1288; CHECK-NEXT:    movl %edx, %ebx
1289; CHECK-NEXT:    movl %esi, %r13d
1290; CHECK-NEXT:    movl %edi, %r15d
1291; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1292; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1293; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1294; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1295; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1296; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1297; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1298; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1299; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1300; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1301; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %eax
1302; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1303; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %r14d
1304; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %ebp
1305; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %r12d
1306; CHECK-NEXT:    movzwl {{[0-9]+}}(%rsp), %edi
1307; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1308; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1309; CHECK-NEXT:    movzwl %r15w, %edi
1310; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1311; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1312; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1313; CHECK-NEXT:    callq fmodf@PLT
1314; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1315; CHECK-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
1316; CHECK-NEXT:    movl %r12d, %edi
1317; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1318; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1319; CHECK-NEXT:    movzwl %r13w, %edi
1320; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1321; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1322; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1323; CHECK-NEXT:    callq fmodf@PLT
1324; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1325; CHECK-NEXT:    movl %eax, %r12d
1326; CHECK-NEXT:    movl %ebp, %edi
1327; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1328; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1329; CHECK-NEXT:    movzwl %bx, %edi
1330; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1331; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1332; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1333; CHECK-NEXT:    callq fmodf@PLT
1334; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1335; CHECK-NEXT:    movl %eax, %r13d
1336; CHECK-NEXT:    movl %r14d, %edi
1337; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1338; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1339; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 2-byte Folded Reload
1340; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1341; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1342; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1343; CHECK-NEXT:    callq fmodf@PLT
1344; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1345; CHECK-NEXT:    movl %eax, %r14d
1346; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1347; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1348; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1349; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 2-byte Folded Reload
1350; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1351; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1352; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1353; CHECK-NEXT:    callq fmodf@PLT
1354; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1355; CHECK-NEXT:    movl %eax, %r15d
1356; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1357; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1358; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1359; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 2-byte Folded Reload
1360; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1361; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1362; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1363; CHECK-NEXT:    callq fmodf@PLT
1364; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1365; CHECK-NEXT:    movl %eax, %ebp
1366; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1367; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1368; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1369; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1370; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1371; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1372; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1373; CHECK-NEXT:    callq fmodf@PLT
1374; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1375; CHECK-NEXT:    movl %eax, %ebx
1376; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1377; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1378; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1379; CHECK-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Reload
1380; CHECK-NEXT:    callq __gnu_h2f_ieee@PLT
1381; CHECK-NEXT:    movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Reload
1382; CHECK-NEXT:    # xmm1 = mem[0],zero,zero,zero
1383; CHECK-NEXT:    callq fmodf@PLT
1384; CHECK-NEXT:    callq __gnu_f2h_ieee@PLT
1385; CHECK-NEXT:    movq {{[0-9]+}}(%rsp), %rcx
1386; CHECK-NEXT:    movw %ax, 14(%rcx)
1387; CHECK-NEXT:    movw %bx, 12(%rcx)
1388; CHECK-NEXT:    movw %bp, 10(%rcx)
1389; CHECK-NEXT:    movw %r15w, 8(%rcx)
1390; CHECK-NEXT:    movw %r14w, 6(%rcx)
1391; CHECK-NEXT:    movw %r13w, 4(%rcx)
1392; CHECK-NEXT:    movw %r12w, 2(%rcx)
1393; CHECK-NEXT:    movzwl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 2-byte Folded Reload
1394; CHECK-NEXT:    movw %ax, (%rcx)
1395; CHECK-NEXT:    addq $56, %rsp
1396; CHECK-NEXT:    popq %rbx
1397; CHECK-NEXT:    popq %r12
1398; CHECK-NEXT:    popq %r13
1399; CHECK-NEXT:    popq %r14
1400; CHECK-NEXT:    popq %r15
1401; CHECK-NEXT:    popq %rbp
1402; CHECK-NEXT:    retq
1403  %frem = frem <8 x half> %a0, %a1
1404  store <8 x half> %frem, <8 x half> *%p3
1405  ret void
1406}
1407
1408define void @frem_v4f80(<4 x x86_fp80> %a0, <4 x x86_fp80> %a1, <4 x x86_fp80> *%p3) nounwind {
1409; CHECK-LABEL: frem_v4f80:
1410; CHECK:       # %bb.0:
1411; CHECK-NEXT:    pushq %rbx
1412; CHECK-NEXT:    subq $128, %rsp
1413; CHECK-NEXT:    movq %rdi, %rbx
1414; CHECK-NEXT:    fldt {{[0-9]+}}(%rsp)
1415; CHECK-NEXT:    fstpt {{[-0-9]+}}(%r{{[sb]}}p) # 10-byte Folded Spill
1416; CHECK-NEXT:    fldt {{[0-9]+}}(%rsp)
1417; CHECK-NEXT:    fstpt {{[-0-9]+}}(%r{{[sb]}}p) # 10-byte Folded Spill
1418; CHECK-NEXT:    fldt {{[0-9]+}}(%rsp)
1419; CHECK-NEXT:    fstpt {{[-0-9]+}}(%r{{[sb]}}p) # 10-byte Folded Spill
1420; CHECK-NEXT:    fldt {{[0-9]+}}(%rsp)
1421; CHECK-NEXT:    fstpt {{[-0-9]+}}(%r{{[sb]}}p) # 10-byte Folded Spill
1422; CHECK-NEXT:    fldt {{[0-9]+}}(%rsp)
1423; CHECK-NEXT:    fstpt {{[-0-9]+}}(%r{{[sb]}}p) # 10-byte Folded Spill
1424; CHECK-NEXT:    fldt {{[0-9]+}}(%rsp)
1425; CHECK-NEXT:    fstpt {{[-0-9]+}}(%r{{[sb]}}p) # 10-byte Folded Spill
1426; CHECK-NEXT:    fldt {{[0-9]+}}(%rsp)
1427; CHECK-NEXT:    fldt {{[0-9]+}}(%rsp)
1428; CHECK-NEXT:    fstpt {{[0-9]+}}(%rsp)
1429; CHECK-NEXT:    fstpt (%rsp)
1430; CHECK-NEXT:    callq fmodl@PLT
1431; CHECK-NEXT:    fstpt {{[-0-9]+}}(%r{{[sb]}}p) # 10-byte Folded Spill
1432; CHECK-NEXT:    fldt {{[-0-9]+}}(%r{{[sb]}}p) # 10-byte Folded Reload
1433; CHECK-NEXT:    fstpt {{[0-9]+}}(%rsp)
1434; CHECK-NEXT:    fldt {{[-0-9]+}}(%r{{[sb]}}p) # 10-byte Folded Reload
1435; CHECK-NEXT:    fstpt (%rsp)
1436; CHECK-NEXT:    callq fmodl@PLT
1437; CHECK-NEXT:    fstpt {{[-0-9]+}}(%r{{[sb]}}p) # 10-byte Folded Spill
1438; CHECK-NEXT:    fldt {{[-0-9]+}}(%r{{[sb]}}p) # 10-byte Folded Reload
1439; CHECK-NEXT:    fstpt {{[0-9]+}}(%rsp)
1440; CHECK-NEXT:    fldt {{[-0-9]+}}(%r{{[sb]}}p) # 10-byte Folded Reload
1441; CHECK-NEXT:    fstpt (%rsp)
1442; CHECK-NEXT:    callq fmodl@PLT
1443; CHECK-NEXT:    fstpt {{[-0-9]+}}(%r{{[sb]}}p) # 10-byte Folded Spill
1444; CHECK-NEXT:    fldt {{[-0-9]+}}(%r{{[sb]}}p) # 10-byte Folded Reload
1445; CHECK-NEXT:    fstpt {{[0-9]+}}(%rsp)
1446; CHECK-NEXT:    fldt {{[-0-9]+}}(%r{{[sb]}}p) # 10-byte Folded Reload
1447; CHECK-NEXT:    fstpt (%rsp)
1448; CHECK-NEXT:    callq fmodl@PLT
1449; CHECK-NEXT:    fstpt 30(%rbx)
1450; CHECK-NEXT:    fldt {{[-0-9]+}}(%r{{[sb]}}p) # 10-byte Folded Reload
1451; CHECK-NEXT:    fstpt 20(%rbx)
1452; CHECK-NEXT:    fldt {{[-0-9]+}}(%r{{[sb]}}p) # 10-byte Folded Reload
1453; CHECK-NEXT:    fstpt 10(%rbx)
1454; CHECK-NEXT:    fldt {{[-0-9]+}}(%r{{[sb]}}p) # 10-byte Folded Reload
1455; CHECK-NEXT:    fstpt (%rbx)
1456; CHECK-NEXT:    addq $128, %rsp
1457; CHECK-NEXT:    popq %rbx
1458; CHECK-NEXT:    retq
1459  %frem = frem <4 x x86_fp80> %a0, %a1
1460  store <4 x x86_fp80> %frem, <4 x x86_fp80> *%p3
1461  ret void
1462}
1463