xref: /llvm-project/llvm/test/CodeGen/X86/recip-fastmath.ll (revision a2a0089ac3a5781ba74d4d319c87c9e8b46d4eda) 
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2     | FileCheck %s --check-prefix=SSE
3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx      | FileCheck %s --check-prefixes=AVX,AVX-RECIP
4; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx,+fma | FileCheck %s --check-prefixes=AVX,FMA-RECIP
5; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=bdver2 | FileCheck %s --check-prefixes=AVX,BDVER2
6; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=btver2 | FileCheck %s --check-prefixes=AVX,BTVER2
7; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=sandybridge | FileCheck %s --check-prefixes=AVX,SANDY
8; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=haswell | FileCheck %s --check-prefixes=AVX,HASWELL
9; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=haswell -mattr=-fma | FileCheck %s --check-prefixes=AVX,HASWELL-NO-FMA
10; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=knl | FileCheck %s --check-prefixes=AVX,AVX512,KNL
11; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=skx | FileCheck %s --check-prefixes=AVX,AVX512,SKX
12
13; If the target's divss/divps instructions are substantially
14; slower than rcpss/rcpps with a Newton-Raphson refinement,
15; we should generate the estimate sequence.
16
17; See PR21385 ( http://llvm.org/bugs/show_bug.cgi?id=21385 )
18; for details about the accuracy, speed, and implementation
19; differences of x86 reciprocal estimates.
20
21define float @f32_no_estimate(float %x) #0 {
22; SSE-LABEL: f32_no_estimate:
23; SSE:       # %bb.0:
24; SSE-NEXT:    movss {{.*#+}} xmm1 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
25; SSE-NEXT:    divss %xmm0, %xmm1
26; SSE-NEXT:    movaps %xmm1, %xmm0
27; SSE-NEXT:    retq
28;
29; AVX-LABEL: f32_no_estimate:
30; AVX:       # %bb.0:
31; AVX-NEXT:    vmovss {{.*#+}} xmm1 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
32; AVX-NEXT:    vdivss %xmm0, %xmm1, %xmm0
33; AVX-NEXT:    retq
34  %div = fdiv fast float 1.0, %x
35  ret float %div
36}
37
38define float @f32_one_step(float %x) #1 {
39; SSE-LABEL: f32_one_step:
40; SSE:       # %bb.0:
41; SSE-NEXT:    rcpss %xmm0, %xmm2
42; SSE-NEXT:    mulss %xmm2, %xmm0
43; SSE-NEXT:    movss {{.*#+}} xmm1 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
44; SSE-NEXT:    subss %xmm0, %xmm1
45; SSE-NEXT:    mulss %xmm2, %xmm1
46; SSE-NEXT:    addss %xmm2, %xmm1
47; SSE-NEXT:    movaps %xmm1, %xmm0
48; SSE-NEXT:    retq
49;
50; AVX-RECIP-LABEL: f32_one_step:
51; AVX-RECIP:       # %bb.0:
52; AVX-RECIP-NEXT:    vrcpss %xmm0, %xmm0, %xmm1
53; AVX-RECIP-NEXT:    vmulss %xmm1, %xmm0, %xmm0
54; AVX-RECIP-NEXT:    vmovss {{.*#+}} xmm2 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
55; AVX-RECIP-NEXT:    vsubss %xmm0, %xmm2, %xmm0
56; AVX-RECIP-NEXT:    vmulss %xmm0, %xmm1, %xmm0
57; AVX-RECIP-NEXT:    vaddss %xmm0, %xmm1, %xmm0
58; AVX-RECIP-NEXT:    retq
59;
60; FMA-RECIP-LABEL: f32_one_step:
61; FMA-RECIP:       # %bb.0:
62; FMA-RECIP-NEXT:    vrcpss %xmm0, %xmm0, %xmm1
63; FMA-RECIP-NEXT:    vfmsub213ss {{.*#+}} xmm0 = (xmm1 * xmm0) - mem
64; FMA-RECIP-NEXT:    vfnmadd132ss {{.*#+}} xmm0 = -(xmm0 * xmm1) + xmm1
65; FMA-RECIP-NEXT:    retq
66;
67; BDVER2-LABEL: f32_one_step:
68; BDVER2:       # %bb.0:
69; BDVER2-NEXT:    vrcpss %xmm0, %xmm0, %xmm1
70; BDVER2-NEXT:    vfmsubss {{.*#+}} xmm0 = (xmm0 * xmm1) - mem
71; BDVER2-NEXT:    vfnmaddss {{.*#+}} xmm0 = -(xmm1 * xmm0) + xmm1
72; BDVER2-NEXT:    retq
73;
74; BTVER2-LABEL: f32_one_step:
75; BTVER2:       # %bb.0:
76; BTVER2-NEXT:    vmovss {{.*#+}} xmm2 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
77; BTVER2-NEXT:    vrcpss %xmm0, %xmm0, %xmm1
78; BTVER2-NEXT:    vmulss %xmm1, %xmm0, %xmm0
79; BTVER2-NEXT:    vsubss %xmm0, %xmm2, %xmm0
80; BTVER2-NEXT:    vmulss %xmm0, %xmm1, %xmm0
81; BTVER2-NEXT:    vaddss %xmm0, %xmm1, %xmm0
82; BTVER2-NEXT:    retq
83;
84; SANDY-LABEL: f32_one_step:
85; SANDY:       # %bb.0:
86; SANDY-NEXT:    vrcpss %xmm0, %xmm0, %xmm1
87; SANDY-NEXT:    vmulss %xmm1, %xmm0, %xmm0
88; SANDY-NEXT:    vmovss {{.*#+}} xmm2 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
89; SANDY-NEXT:    vsubss %xmm0, %xmm2, %xmm0
90; SANDY-NEXT:    vmulss %xmm0, %xmm1, %xmm0
91; SANDY-NEXT:    vaddss %xmm0, %xmm1, %xmm0
92; SANDY-NEXT:    retq
93;
94; HASWELL-LABEL: f32_one_step:
95; HASWELL:       # %bb.0:
96; HASWELL-NEXT:    vrcpss %xmm0, %xmm0, %xmm1
97; HASWELL-NEXT:    vfmsub213ss {{.*#+}} xmm0 = (xmm1 * xmm0) - mem
98; HASWELL-NEXT:    vfnmadd132ss {{.*#+}} xmm0 = -(xmm0 * xmm1) + xmm1
99; HASWELL-NEXT:    retq
100;
101; HASWELL-NO-FMA-LABEL: f32_one_step:
102; HASWELL-NO-FMA:       # %bb.0:
103; HASWELL-NO-FMA-NEXT:    vrcpss %xmm0, %xmm0, %xmm1
104; HASWELL-NO-FMA-NEXT:    vmulss %xmm1, %xmm0, %xmm0
105; HASWELL-NO-FMA-NEXT:    vmovss {{.*#+}} xmm2 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
106; HASWELL-NO-FMA-NEXT:    vsubss %xmm0, %xmm2, %xmm0
107; HASWELL-NO-FMA-NEXT:    vmulss %xmm0, %xmm1, %xmm0
108; HASWELL-NO-FMA-NEXT:    vaddss %xmm0, %xmm1, %xmm0
109; HASWELL-NO-FMA-NEXT:    retq
110;
111; AVX512-LABEL: f32_one_step:
112; AVX512:       # %bb.0:
113; AVX512-NEXT:    vrcpss %xmm0, %xmm0, %xmm1
114; AVX512-NEXT:    vfmsub213ss {{.*#+}} xmm0 = (xmm1 * xmm0) - mem
115; AVX512-NEXT:    vfnmadd132ss {{.*#+}} xmm0 = -(xmm0 * xmm1) + xmm1
116; AVX512-NEXT:    retq
117  %div = fdiv fast float 1.0, %x
118  ret float %div
119}
120
121define float @f32_one_step_variables(float %x, float %y) #1 {
122; SSE-LABEL: f32_one_step_variables:
123; SSE:       # %bb.0:
124; SSE-NEXT:    rcpss %xmm1, %xmm2
125; SSE-NEXT:    movaps %xmm0, %xmm3
126; SSE-NEXT:    mulss %xmm2, %xmm3
127; SSE-NEXT:    mulss %xmm3, %xmm1
128; SSE-NEXT:    subss %xmm1, %xmm0
129; SSE-NEXT:    mulss %xmm2, %xmm0
130; SSE-NEXT:    addss %xmm3, %xmm0
131; SSE-NEXT:    retq
132;
133; AVX-RECIP-LABEL: f32_one_step_variables:
134; AVX-RECIP:       # %bb.0:
135; AVX-RECIP-NEXT:    vrcpss %xmm1, %xmm1, %xmm2
136; AVX-RECIP-NEXT:    vmulss %xmm2, %xmm0, %xmm3
137; AVX-RECIP-NEXT:    vmulss %xmm3, %xmm1, %xmm1
138; AVX-RECIP-NEXT:    vsubss %xmm1, %xmm0, %xmm0
139; AVX-RECIP-NEXT:    vmulss %xmm0, %xmm2, %xmm0
140; AVX-RECIP-NEXT:    vaddss %xmm0, %xmm3, %xmm0
141; AVX-RECIP-NEXT:    retq
142;
143; FMA-RECIP-LABEL: f32_one_step_variables:
144; FMA-RECIP:       # %bb.0:
145; FMA-RECIP-NEXT:    vrcpss %xmm1, %xmm1, %xmm2
146; FMA-RECIP-NEXT:    vmulss %xmm2, %xmm0, %xmm3
147; FMA-RECIP-NEXT:    vfmsub231ss {{.*#+}} xmm0 = (xmm3 * xmm1) - xmm0
148; FMA-RECIP-NEXT:    vfnmadd213ss {{.*#+}} xmm0 = -(xmm2 * xmm0) + xmm3
149; FMA-RECIP-NEXT:    retq
150;
151; BDVER2-LABEL: f32_one_step_variables:
152; BDVER2:       # %bb.0:
153; BDVER2-NEXT:    vrcpss %xmm1, %xmm1, %xmm2
154; BDVER2-NEXT:    vmulss %xmm2, %xmm0, %xmm3
155; BDVER2-NEXT:    vfmsubss {{.*#+}} xmm0 = (xmm1 * xmm3) - xmm0
156; BDVER2-NEXT:    vfnmaddss {{.*#+}} xmm0 = -(xmm2 * xmm0) + xmm3
157; BDVER2-NEXT:    retq
158;
159; BTVER2-LABEL: f32_one_step_variables:
160; BTVER2:       # %bb.0:
161; BTVER2-NEXT:    vrcpss %xmm1, %xmm1, %xmm2
162; BTVER2-NEXT:    vmulss %xmm2, %xmm0, %xmm3
163; BTVER2-NEXT:    vmulss %xmm3, %xmm1, %xmm1
164; BTVER2-NEXT:    vsubss %xmm1, %xmm0, %xmm0
165; BTVER2-NEXT:    vmulss %xmm0, %xmm2, %xmm0
166; BTVER2-NEXT:    vaddss %xmm0, %xmm3, %xmm0
167; BTVER2-NEXT:    retq
168;
169; SANDY-LABEL: f32_one_step_variables:
170; SANDY:       # %bb.0:
171; SANDY-NEXT:    vrcpss %xmm1, %xmm1, %xmm2
172; SANDY-NEXT:    vmulss %xmm2, %xmm0, %xmm3
173; SANDY-NEXT:    vmulss %xmm3, %xmm1, %xmm1
174; SANDY-NEXT:    vsubss %xmm1, %xmm0, %xmm0
175; SANDY-NEXT:    vmulss %xmm0, %xmm2, %xmm0
176; SANDY-NEXT:    vaddss %xmm0, %xmm3, %xmm0
177; SANDY-NEXT:    retq
178;
179; HASWELL-LABEL: f32_one_step_variables:
180; HASWELL:       # %bb.0:
181; HASWELL-NEXT:    vrcpss %xmm1, %xmm1, %xmm2
182; HASWELL-NEXT:    vmulss %xmm2, %xmm0, %xmm3
183; HASWELL-NEXT:    vfmsub231ss {{.*#+}} xmm0 = (xmm3 * xmm1) - xmm0
184; HASWELL-NEXT:    vfnmadd213ss {{.*#+}} xmm0 = -(xmm2 * xmm0) + xmm3
185; HASWELL-NEXT:    retq
186;
187; HASWELL-NO-FMA-LABEL: f32_one_step_variables:
188; HASWELL-NO-FMA:       # %bb.0:
189; HASWELL-NO-FMA-NEXT:    vrcpss %xmm1, %xmm1, %xmm2
190; HASWELL-NO-FMA-NEXT:    vmulss %xmm2, %xmm0, %xmm3
191; HASWELL-NO-FMA-NEXT:    vmulss %xmm3, %xmm1, %xmm1
192; HASWELL-NO-FMA-NEXT:    vsubss %xmm1, %xmm0, %xmm0
193; HASWELL-NO-FMA-NEXT:    vmulss %xmm0, %xmm2, %xmm0
194; HASWELL-NO-FMA-NEXT:    vaddss %xmm0, %xmm3, %xmm0
195; HASWELL-NO-FMA-NEXT:    retq
196;
197; AVX512-LABEL: f32_one_step_variables:
198; AVX512:       # %bb.0:
199; AVX512-NEXT:    vrcpss %xmm1, %xmm1, %xmm2
200; AVX512-NEXT:    vmulss %xmm2, %xmm0, %xmm3
201; AVX512-NEXT:    vfmsub231ss {{.*#+}} xmm0 = (xmm3 * xmm1) - xmm0
202; AVX512-NEXT:    vfnmadd213ss {{.*#+}} xmm0 = -(xmm2 * xmm0) + xmm3
203; AVX512-NEXT:    retq
204  %div = fdiv fast float %x, %y
205  ret float %div
206}
207
208define float @f32_two_step(float %x) #2 {
209; SSE-LABEL: f32_two_step:
210; SSE:       # %bb.0:
211; SSE-NEXT:    rcpss %xmm0, %xmm2
212; SSE-NEXT:    movaps %xmm0, %xmm3
213; SSE-NEXT:    mulss %xmm2, %xmm3
214; SSE-NEXT:    movss {{.*#+}} xmm1 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
215; SSE-NEXT:    movaps %xmm1, %xmm4
216; SSE-NEXT:    subss %xmm3, %xmm4
217; SSE-NEXT:    mulss %xmm2, %xmm4
218; SSE-NEXT:    addss %xmm2, %xmm4
219; SSE-NEXT:    mulss %xmm4, %xmm0
220; SSE-NEXT:    subss %xmm0, %xmm1
221; SSE-NEXT:    mulss %xmm4, %xmm1
222; SSE-NEXT:    addss %xmm4, %xmm1
223; SSE-NEXT:    movaps %xmm1, %xmm0
224; SSE-NEXT:    retq
225;
226; AVX-RECIP-LABEL: f32_two_step:
227; AVX-RECIP:       # %bb.0:
228; AVX-RECIP-NEXT:    vrcpss %xmm0, %xmm0, %xmm1
229; AVX-RECIP-NEXT:    vmulss %xmm1, %xmm0, %xmm2
230; AVX-RECIP-NEXT:    vmovss {{.*#+}} xmm3 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
231; AVX-RECIP-NEXT:    vsubss %xmm2, %xmm3, %xmm2
232; AVX-RECIP-NEXT:    vmulss %xmm2, %xmm1, %xmm2
233; AVX-RECIP-NEXT:    vaddss %xmm2, %xmm1, %xmm1
234; AVX-RECIP-NEXT:    vmulss %xmm1, %xmm0, %xmm0
235; AVX-RECIP-NEXT:    vsubss %xmm0, %xmm3, %xmm0
236; AVX-RECIP-NEXT:    vmulss %xmm0, %xmm1, %xmm0
237; AVX-RECIP-NEXT:    vaddss %xmm0, %xmm1, %xmm0
238; AVX-RECIP-NEXT:    retq
239;
240; FMA-RECIP-LABEL: f32_two_step:
241; FMA-RECIP:       # %bb.0:
242; FMA-RECIP-NEXT:    vrcpss %xmm0, %xmm0, %xmm1
243; FMA-RECIP-NEXT:    vmovss {{.*#+}} xmm2 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
244; FMA-RECIP-NEXT:    vmovaps %xmm1, %xmm3
245; FMA-RECIP-NEXT:    vfmsub213ss {{.*#+}} xmm3 = (xmm0 * xmm3) - xmm2
246; FMA-RECIP-NEXT:    vfnmadd132ss {{.*#+}} xmm3 = -(xmm3 * xmm1) + xmm1
247; FMA-RECIP-NEXT:    vfmsub213ss {{.*#+}} xmm0 = (xmm3 * xmm0) - xmm2
248; FMA-RECIP-NEXT:    vfnmadd132ss {{.*#+}} xmm0 = -(xmm0 * xmm3) + xmm3
249; FMA-RECIP-NEXT:    retq
250;
251; BDVER2-LABEL: f32_two_step:
252; BDVER2:       # %bb.0:
253; BDVER2-NEXT:    vrcpss %xmm0, %xmm0, %xmm1
254; BDVER2-NEXT:    vmovss {{.*#+}} xmm2 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
255; BDVER2-NEXT:    vfmsubss {{.*#+}} xmm3 = (xmm0 * xmm1) - xmm2
256; BDVER2-NEXT:    vfnmaddss {{.*#+}} xmm1 = -(xmm1 * xmm3) + xmm1
257; BDVER2-NEXT:    vfmsubss {{.*#+}} xmm0 = (xmm0 * xmm1) - xmm2
258; BDVER2-NEXT:    vfnmaddss {{.*#+}} xmm0 = -(xmm1 * xmm0) + xmm1
259; BDVER2-NEXT:    retq
260;
261; BTVER2-LABEL: f32_two_step:
262; BTVER2:       # %bb.0:
263; BTVER2-NEXT:    vmovss {{.*#+}} xmm3 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
264; BTVER2-NEXT:    vrcpss %xmm0, %xmm0, %xmm1
265; BTVER2-NEXT:    vmulss %xmm1, %xmm0, %xmm2
266; BTVER2-NEXT:    vsubss %xmm2, %xmm3, %xmm2
267; BTVER2-NEXT:    vmulss %xmm2, %xmm1, %xmm2
268; BTVER2-NEXT:    vaddss %xmm2, %xmm1, %xmm1
269; BTVER2-NEXT:    vmulss %xmm1, %xmm0, %xmm0
270; BTVER2-NEXT:    vsubss %xmm0, %xmm3, %xmm0
271; BTVER2-NEXT:    vmulss %xmm0, %xmm1, %xmm0
272; BTVER2-NEXT:    vaddss %xmm0, %xmm1, %xmm0
273; BTVER2-NEXT:    retq
274;
275; SANDY-LABEL: f32_two_step:
276; SANDY:       # %bb.0:
277; SANDY-NEXT:    vrcpss %xmm0, %xmm0, %xmm1
278; SANDY-NEXT:    vmulss %xmm1, %xmm0, %xmm2
279; SANDY-NEXT:    vmovss {{.*#+}} xmm3 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
280; SANDY-NEXT:    vsubss %xmm2, %xmm3, %xmm2
281; SANDY-NEXT:    vmulss %xmm2, %xmm1, %xmm2
282; SANDY-NEXT:    vaddss %xmm2, %xmm1, %xmm1
283; SANDY-NEXT:    vmulss %xmm1, %xmm0, %xmm0
284; SANDY-NEXT:    vsubss %xmm0, %xmm3, %xmm0
285; SANDY-NEXT:    vmulss %xmm0, %xmm1, %xmm0
286; SANDY-NEXT:    vaddss %xmm0, %xmm1, %xmm0
287; SANDY-NEXT:    retq
288;
289; HASWELL-LABEL: f32_two_step:
290; HASWELL:       # %bb.0:
291; HASWELL-NEXT:    vrcpss %xmm0, %xmm0, %xmm1
292; HASWELL-NEXT:    vmovss {{.*#+}} xmm2 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
293; HASWELL-NEXT:    vmovaps %xmm1, %xmm3
294; HASWELL-NEXT:    vfmsub213ss {{.*#+}} xmm3 = (xmm0 * xmm3) - xmm2
295; HASWELL-NEXT:    vfnmadd132ss {{.*#+}} xmm3 = -(xmm3 * xmm1) + xmm1
296; HASWELL-NEXT:    vfmsub213ss {{.*#+}} xmm0 = (xmm3 * xmm0) - xmm2
297; HASWELL-NEXT:    vfnmadd132ss {{.*#+}} xmm0 = -(xmm0 * xmm3) + xmm3
298; HASWELL-NEXT:    retq
299;
300; HASWELL-NO-FMA-LABEL: f32_two_step:
301; HASWELL-NO-FMA:       # %bb.0:
302; HASWELL-NO-FMA-NEXT:    vrcpss %xmm0, %xmm0, %xmm1
303; HASWELL-NO-FMA-NEXT:    vmulss %xmm1, %xmm0, %xmm2
304; HASWELL-NO-FMA-NEXT:    vmovss {{.*#+}} xmm3 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
305; HASWELL-NO-FMA-NEXT:    vsubss %xmm2, %xmm3, %xmm2
306; HASWELL-NO-FMA-NEXT:    vmulss %xmm2, %xmm1, %xmm2
307; HASWELL-NO-FMA-NEXT:    vaddss %xmm2, %xmm1, %xmm1
308; HASWELL-NO-FMA-NEXT:    vmulss %xmm1, %xmm0, %xmm0
309; HASWELL-NO-FMA-NEXT:    vsubss %xmm0, %xmm3, %xmm0
310; HASWELL-NO-FMA-NEXT:    vmulss %xmm0, %xmm1, %xmm0
311; HASWELL-NO-FMA-NEXT:    vaddss %xmm0, %xmm1, %xmm0
312; HASWELL-NO-FMA-NEXT:    retq
313;
314; AVX512-LABEL: f32_two_step:
315; AVX512:       # %bb.0:
316; AVX512-NEXT:    vrcpss %xmm0, %xmm0, %xmm1
317; AVX512-NEXT:    vmovss {{.*#+}} xmm2 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]
318; AVX512-NEXT:    vmovaps %xmm1, %xmm3
319; AVX512-NEXT:    vfmsub213ss {{.*#+}} xmm3 = (xmm0 * xmm3) - xmm2
320; AVX512-NEXT:    vfnmadd132ss {{.*#+}} xmm3 = -(xmm3 * xmm1) + xmm1
321; AVX512-NEXT:    vfmsub213ss {{.*#+}} xmm0 = (xmm3 * xmm0) - xmm2
322; AVX512-NEXT:    vfnmadd132ss {{.*#+}} xmm0 = -(xmm0 * xmm3) + xmm3
323; AVX512-NEXT:    retq
324  %div = fdiv fast float 1.0, %x
325  ret float %div
326}
327
328define <4 x float> @v4f32_no_estimate(<4 x float> %x) #0 {
329; SSE-LABEL: v4f32_no_estimate:
330; SSE:       # %bb.0:
331; SSE-NEXT:    movaps {{.*#+}} xmm1 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
332; SSE-NEXT:    divps %xmm0, %xmm1
333; SSE-NEXT:    movaps %xmm1, %xmm0
334; SSE-NEXT:    retq
335;
336; AVX-LABEL: v4f32_no_estimate:
337; AVX:       # %bb.0:
338; AVX-NEXT:    vbroadcastss {{.*#+}} xmm1 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
339; AVX-NEXT:    vdivps %xmm0, %xmm1, %xmm0
340; AVX-NEXT:    retq
341  %div = fdiv fast <4 x float> <float 1.0, float 1.0, float 1.0, float 1.0>, %x
342  ret <4 x float> %div
343}
344
345define <4 x float> @v4f32_one_step(<4 x float> %x) #1 {
346; SSE-LABEL: v4f32_one_step:
347; SSE:       # %bb.0:
348; SSE-NEXT:    rcpps %xmm0, %xmm2
349; SSE-NEXT:    mulps %xmm2, %xmm0
350; SSE-NEXT:    movaps {{.*#+}} xmm1 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
351; SSE-NEXT:    subps %xmm0, %xmm1
352; SSE-NEXT:    mulps %xmm2, %xmm1
353; SSE-NEXT:    addps %xmm2, %xmm1
354; SSE-NEXT:    movaps %xmm1, %xmm0
355; SSE-NEXT:    retq
356;
357; AVX-RECIP-LABEL: v4f32_one_step:
358; AVX-RECIP:       # %bb.0:
359; AVX-RECIP-NEXT:    vrcpps %xmm0, %xmm1
360; AVX-RECIP-NEXT:    vmulps %xmm1, %xmm0, %xmm0
361; AVX-RECIP-NEXT:    vbroadcastss {{.*#+}} xmm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
362; AVX-RECIP-NEXT:    vsubps %xmm0, %xmm2, %xmm0
363; AVX-RECIP-NEXT:    vmulps %xmm0, %xmm1, %xmm0
364; AVX-RECIP-NEXT:    vaddps %xmm0, %xmm1, %xmm0
365; AVX-RECIP-NEXT:    retq
366;
367; FMA-RECIP-LABEL: v4f32_one_step:
368; FMA-RECIP:       # %bb.0:
369; FMA-RECIP-NEXT:    vrcpps %xmm0, %xmm1
370; FMA-RECIP-NEXT:    vfmsub213ps {{.*#+}} xmm0 = (xmm1 * xmm0) - mem
371; FMA-RECIP-NEXT:    vfnmadd132ps {{.*#+}} xmm0 = -(xmm0 * xmm1) + xmm1
372; FMA-RECIP-NEXT:    retq
373;
374; BDVER2-LABEL: v4f32_one_step:
375; BDVER2:       # %bb.0:
376; BDVER2-NEXT:    vrcpps %xmm0, %xmm1
377; BDVER2-NEXT:    vfmsubps {{.*#+}} xmm0 = (xmm0 * xmm1) - mem
378; BDVER2-NEXT:    vfnmaddps {{.*#+}} xmm0 = -(xmm1 * xmm0) + xmm1
379; BDVER2-NEXT:    retq
380;
381; BTVER2-LABEL: v4f32_one_step:
382; BTVER2:       # %bb.0:
383; BTVER2-NEXT:    vbroadcastss {{.*#+}} xmm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
384; BTVER2-NEXT:    vrcpps %xmm0, %xmm1
385; BTVER2-NEXT:    vmulps %xmm1, %xmm0, %xmm0
386; BTVER2-NEXT:    vsubps %xmm0, %xmm2, %xmm0
387; BTVER2-NEXT:    vmulps %xmm0, %xmm1, %xmm0
388; BTVER2-NEXT:    vaddps %xmm0, %xmm1, %xmm0
389; BTVER2-NEXT:    retq
390;
391; SANDY-LABEL: v4f32_one_step:
392; SANDY:       # %bb.0:
393; SANDY-NEXT:    vrcpps %xmm0, %xmm1
394; SANDY-NEXT:    vmulps %xmm1, %xmm0, %xmm0
395; SANDY-NEXT:    vbroadcastss {{.*#+}} xmm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
396; SANDY-NEXT:    vsubps %xmm0, %xmm2, %xmm0
397; SANDY-NEXT:    vmulps %xmm0, %xmm1, %xmm0
398; SANDY-NEXT:    vaddps %xmm0, %xmm1, %xmm0
399; SANDY-NEXT:    retq
400;
401; HASWELL-LABEL: v4f32_one_step:
402; HASWELL:       # %bb.0:
403; HASWELL-NEXT:    vrcpps %xmm0, %xmm1
404; HASWELL-NEXT:    vbroadcastss {{.*#+}} xmm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
405; HASWELL-NEXT:    vfmsub213ps {{.*#+}} xmm0 = (xmm1 * xmm0) - xmm2
406; HASWELL-NEXT:    vfnmadd132ps {{.*#+}} xmm0 = -(xmm0 * xmm1) + xmm1
407; HASWELL-NEXT:    retq
408;
409; HASWELL-NO-FMA-LABEL: v4f32_one_step:
410; HASWELL-NO-FMA:       # %bb.0:
411; HASWELL-NO-FMA-NEXT:    vrcpps %xmm0, %xmm1
412; HASWELL-NO-FMA-NEXT:    vmulps %xmm1, %xmm0, %xmm0
413; HASWELL-NO-FMA-NEXT:    vbroadcastss {{.*#+}} xmm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
414; HASWELL-NO-FMA-NEXT:    vsubps %xmm0, %xmm2, %xmm0
415; HASWELL-NO-FMA-NEXT:    vmulps %xmm0, %xmm1, %xmm0
416; HASWELL-NO-FMA-NEXT:    vaddps %xmm0, %xmm1, %xmm0
417; HASWELL-NO-FMA-NEXT:    retq
418;
419; KNL-LABEL: v4f32_one_step:
420; KNL:       # %bb.0:
421; KNL-NEXT:    vrcpps %xmm0, %xmm1
422; KNL-NEXT:    vbroadcastss {{.*#+}} xmm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
423; KNL-NEXT:    vfmsub213ps {{.*#+}} xmm0 = (xmm1 * xmm0) - xmm2
424; KNL-NEXT:    vfnmadd132ps {{.*#+}} xmm0 = -(xmm0 * xmm1) + xmm1
425; KNL-NEXT:    retq
426;
427; SKX-LABEL: v4f32_one_step:
428; SKX:       # %bb.0:
429; SKX-NEXT:    vrcpps %xmm0, %xmm1
430; SKX-NEXT:    vfmsub213ps {{.*#+}} xmm0 = (xmm1 * xmm0) - mem
431; SKX-NEXT:    vfnmadd132ps {{.*#+}} xmm0 = -(xmm0 * xmm1) + xmm1
432; SKX-NEXT:    retq
433  %div = fdiv fast <4 x float> <float 1.0, float 1.0, float 1.0, float 1.0>, %x
434  ret <4 x float> %div
435}
436
437define <4 x float> @v4f32_one_step_variables(<4 x float> %x, <4 x float> %y) #1 {
438; SSE-LABEL: v4f32_one_step_variables:
439; SSE:       # %bb.0:
440; SSE-NEXT:    rcpps %xmm1, %xmm2
441; SSE-NEXT:    movaps %xmm0, %xmm3
442; SSE-NEXT:    mulps %xmm2, %xmm3
443; SSE-NEXT:    mulps %xmm3, %xmm1
444; SSE-NEXT:    subps %xmm1, %xmm0
445; SSE-NEXT:    mulps %xmm2, %xmm0
446; SSE-NEXT:    addps %xmm3, %xmm0
447; SSE-NEXT:    retq
448;
449; AVX-RECIP-LABEL: v4f32_one_step_variables:
450; AVX-RECIP:       # %bb.0:
451; AVX-RECIP-NEXT:    vrcpps %xmm1, %xmm2
452; AVX-RECIP-NEXT:    vmulps %xmm2, %xmm0, %xmm3
453; AVX-RECIP-NEXT:    vmulps %xmm3, %xmm1, %xmm1
454; AVX-RECIP-NEXT:    vsubps %xmm1, %xmm0, %xmm0
455; AVX-RECIP-NEXT:    vmulps %xmm0, %xmm2, %xmm0
456; AVX-RECIP-NEXT:    vaddps %xmm0, %xmm3, %xmm0
457; AVX-RECIP-NEXT:    retq
458;
459; FMA-RECIP-LABEL: v4f32_one_step_variables:
460; FMA-RECIP:       # %bb.0:
461; FMA-RECIP-NEXT:    vrcpps %xmm1, %xmm2
462; FMA-RECIP-NEXT:    vmulps %xmm2, %xmm0, %xmm3
463; FMA-RECIP-NEXT:    vfmsub231ps {{.*#+}} xmm0 = (xmm3 * xmm1) - xmm0
464; FMA-RECIP-NEXT:    vfnmadd213ps {{.*#+}} xmm0 = -(xmm2 * xmm0) + xmm3
465; FMA-RECIP-NEXT:    retq
466;
467; BDVER2-LABEL: v4f32_one_step_variables:
468; BDVER2:       # %bb.0:
469; BDVER2-NEXT:    vrcpps %xmm1, %xmm2
470; BDVER2-NEXT:    vmulps %xmm2, %xmm0, %xmm3
471; BDVER2-NEXT:    vfmsubps {{.*#+}} xmm0 = (xmm1 * xmm3) - xmm0
472; BDVER2-NEXT:    vfnmaddps {{.*#+}} xmm0 = -(xmm2 * xmm0) + xmm3
473; BDVER2-NEXT:    retq
474;
475; BTVER2-LABEL: v4f32_one_step_variables:
476; BTVER2:       # %bb.0:
477; BTVER2-NEXT:    vrcpps %xmm1, %xmm2
478; BTVER2-NEXT:    vmulps %xmm2, %xmm0, %xmm3
479; BTVER2-NEXT:    vmulps %xmm3, %xmm1, %xmm1
480; BTVER2-NEXT:    vsubps %xmm1, %xmm0, %xmm0
481; BTVER2-NEXT:    vmulps %xmm0, %xmm2, %xmm0
482; BTVER2-NEXT:    vaddps %xmm0, %xmm3, %xmm0
483; BTVER2-NEXT:    retq
484;
485; SANDY-LABEL: v4f32_one_step_variables:
486; SANDY:       # %bb.0:
487; SANDY-NEXT:    vrcpps %xmm1, %xmm2
488; SANDY-NEXT:    vmulps %xmm2, %xmm0, %xmm3
489; SANDY-NEXT:    vmulps %xmm3, %xmm1, %xmm1
490; SANDY-NEXT:    vsubps %xmm1, %xmm0, %xmm0
491; SANDY-NEXT:    vmulps %xmm0, %xmm2, %xmm0
492; SANDY-NEXT:    vaddps %xmm0, %xmm3, %xmm0
493; SANDY-NEXT:    retq
494;
495; HASWELL-LABEL: v4f32_one_step_variables:
496; HASWELL:       # %bb.0:
497; HASWELL-NEXT:    vrcpps %xmm1, %xmm2
498; HASWELL-NEXT:    vmulps %xmm2, %xmm0, %xmm3
499; HASWELL-NEXT:    vfmsub231ps {{.*#+}} xmm0 = (xmm3 * xmm1) - xmm0
500; HASWELL-NEXT:    vfnmadd213ps {{.*#+}} xmm0 = -(xmm2 * xmm0) + xmm3
501; HASWELL-NEXT:    retq
502;
503; HASWELL-NO-FMA-LABEL: v4f32_one_step_variables:
504; HASWELL-NO-FMA:       # %bb.0:
505; HASWELL-NO-FMA-NEXT:    vrcpps %xmm1, %xmm2
506; HASWELL-NO-FMA-NEXT:    vmulps %xmm2, %xmm0, %xmm3
507; HASWELL-NO-FMA-NEXT:    vmulps %xmm3, %xmm1, %xmm1
508; HASWELL-NO-FMA-NEXT:    vsubps %xmm1, %xmm0, %xmm0
509; HASWELL-NO-FMA-NEXT:    vmulps %xmm0, %xmm2, %xmm0
510; HASWELL-NO-FMA-NEXT:    vaddps %xmm0, %xmm3, %xmm0
511; HASWELL-NO-FMA-NEXT:    retq
512;
513; AVX512-LABEL: v4f32_one_step_variables:
514; AVX512:       # %bb.0:
515; AVX512-NEXT:    vrcpps %xmm1, %xmm2
516; AVX512-NEXT:    vmulps %xmm2, %xmm0, %xmm3
517; AVX512-NEXT:    vfmsub231ps {{.*#+}} xmm0 = (xmm3 * xmm1) - xmm0
518; AVX512-NEXT:    vfnmadd213ps {{.*#+}} xmm0 = -(xmm2 * xmm0) + xmm3
519; AVX512-NEXT:    retq
520  %div = fdiv fast <4 x float> %x, %y
521  ret <4 x float> %div
522}
523
524define <4 x float> @v4f32_two_step(<4 x float> %x) #2 {
525; SSE-LABEL: v4f32_two_step:
526; SSE:       # %bb.0:
527; SSE-NEXT:    rcpps %xmm0, %xmm2
528; SSE-NEXT:    movaps %xmm0, %xmm3
529; SSE-NEXT:    mulps %xmm2, %xmm3
530; SSE-NEXT:    movaps {{.*#+}} xmm1 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
531; SSE-NEXT:    movaps %xmm1, %xmm4
532; SSE-NEXT:    subps %xmm3, %xmm4
533; SSE-NEXT:    mulps %xmm2, %xmm4
534; SSE-NEXT:    addps %xmm2, %xmm4
535; SSE-NEXT:    mulps %xmm4, %xmm0
536; SSE-NEXT:    subps %xmm0, %xmm1
537; SSE-NEXT:    mulps %xmm4, %xmm1
538; SSE-NEXT:    addps %xmm4, %xmm1
539; SSE-NEXT:    movaps %xmm1, %xmm0
540; SSE-NEXT:    retq
541;
542; AVX-RECIP-LABEL: v4f32_two_step:
543; AVX-RECIP:       # %bb.0:
544; AVX-RECIP-NEXT:    vrcpps %xmm0, %xmm1
545; AVX-RECIP-NEXT:    vmulps %xmm1, %xmm0, %xmm2
546; AVX-RECIP-NEXT:    vbroadcastss {{.*#+}} xmm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
547; AVX-RECIP-NEXT:    vsubps %xmm2, %xmm3, %xmm2
548; AVX-RECIP-NEXT:    vmulps %xmm2, %xmm1, %xmm2
549; AVX-RECIP-NEXT:    vaddps %xmm2, %xmm1, %xmm1
550; AVX-RECIP-NEXT:    vmulps %xmm1, %xmm0, %xmm0
551; AVX-RECIP-NEXT:    vsubps %xmm0, %xmm3, %xmm0
552; AVX-RECIP-NEXT:    vmulps %xmm0, %xmm1, %xmm0
553; AVX-RECIP-NEXT:    vaddps %xmm0, %xmm1, %xmm0
554; AVX-RECIP-NEXT:    retq
555;
556; FMA-RECIP-LABEL: v4f32_two_step:
557; FMA-RECIP:       # %bb.0:
558; FMA-RECIP-NEXT:    vrcpps %xmm0, %xmm1
559; FMA-RECIP-NEXT:    vbroadcastss {{.*#+}} xmm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
560; FMA-RECIP-NEXT:    vmovaps %xmm1, %xmm3
561; FMA-RECIP-NEXT:    vfmsub213ps {{.*#+}} xmm3 = (xmm0 * xmm3) - xmm2
562; FMA-RECIP-NEXT:    vfnmadd132ps {{.*#+}} xmm3 = -(xmm3 * xmm1) + xmm1
563; FMA-RECIP-NEXT:    vfmsub213ps {{.*#+}} xmm0 = (xmm3 * xmm0) - xmm2
564; FMA-RECIP-NEXT:    vfnmadd132ps {{.*#+}} xmm0 = -(xmm0 * xmm3) + xmm3
565; FMA-RECIP-NEXT:    retq
566;
567; BDVER2-LABEL: v4f32_two_step:
568; BDVER2:       # %bb.0:
569; BDVER2-NEXT:    vrcpps %xmm0, %xmm1
570; BDVER2-NEXT:    vbroadcastss {{.*#+}} xmm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
571; BDVER2-NEXT:    vfmsubps {{.*#+}} xmm3 = (xmm0 * xmm1) - xmm2
572; BDVER2-NEXT:    vfnmaddps {{.*#+}} xmm1 = -(xmm1 * xmm3) + xmm1
573; BDVER2-NEXT:    vfmsubps {{.*#+}} xmm0 = (xmm0 * xmm1) - xmm2
574; BDVER2-NEXT:    vfnmaddps {{.*#+}} xmm0 = -(xmm1 * xmm0) + xmm1
575; BDVER2-NEXT:    retq
576;
577; BTVER2-LABEL: v4f32_two_step:
578; BTVER2:       # %bb.0:
579; BTVER2-NEXT:    vbroadcastss {{.*#+}} xmm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
580; BTVER2-NEXT:    vrcpps %xmm0, %xmm1
581; BTVER2-NEXT:    vmulps %xmm1, %xmm0, %xmm2
582; BTVER2-NEXT:    vsubps %xmm2, %xmm3, %xmm2
583; BTVER2-NEXT:    vmulps %xmm2, %xmm1, %xmm2
584; BTVER2-NEXT:    vaddps %xmm2, %xmm1, %xmm1
585; BTVER2-NEXT:    vmulps %xmm1, %xmm0, %xmm0
586; BTVER2-NEXT:    vsubps %xmm0, %xmm3, %xmm0
587; BTVER2-NEXT:    vmulps %xmm0, %xmm1, %xmm0
588; BTVER2-NEXT:    vaddps %xmm0, %xmm1, %xmm0
589; BTVER2-NEXT:    retq
590;
591; SANDY-LABEL: v4f32_two_step:
592; SANDY:       # %bb.0:
593; SANDY-NEXT:    vrcpps %xmm0, %xmm1
594; SANDY-NEXT:    vmulps %xmm1, %xmm0, %xmm2
595; SANDY-NEXT:    vbroadcastss {{.*#+}} xmm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
596; SANDY-NEXT:    vsubps %xmm2, %xmm3, %xmm2
597; SANDY-NEXT:    vmulps %xmm2, %xmm1, %xmm2
598; SANDY-NEXT:    vaddps %xmm2, %xmm1, %xmm1
599; SANDY-NEXT:    vmulps %xmm1, %xmm0, %xmm0
600; SANDY-NEXT:    vsubps %xmm0, %xmm3, %xmm0
601; SANDY-NEXT:    vmulps %xmm0, %xmm1, %xmm0
602; SANDY-NEXT:    vaddps %xmm0, %xmm1, %xmm0
603; SANDY-NEXT:    retq
604;
605; HASWELL-LABEL: v4f32_two_step:
606; HASWELL:       # %bb.0:
607; HASWELL-NEXT:    vrcpps %xmm0, %xmm1
608; HASWELL-NEXT:    vbroadcastss {{.*#+}} xmm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
609; HASWELL-NEXT:    vmovaps %xmm1, %xmm3
610; HASWELL-NEXT:    vfmsub213ps {{.*#+}} xmm3 = (xmm0 * xmm3) - xmm2
611; HASWELL-NEXT:    vfnmadd132ps {{.*#+}} xmm3 = -(xmm3 * xmm1) + xmm1
612; HASWELL-NEXT:    vfmsub213ps {{.*#+}} xmm0 = (xmm3 * xmm0) - xmm2
613; HASWELL-NEXT:    vfnmadd132ps {{.*#+}} xmm0 = -(xmm0 * xmm3) + xmm3
614; HASWELL-NEXT:    retq
615;
616; HASWELL-NO-FMA-LABEL: v4f32_two_step:
617; HASWELL-NO-FMA:       # %bb.0:
618; HASWELL-NO-FMA-NEXT:    vrcpps %xmm0, %xmm1
619; HASWELL-NO-FMA-NEXT:    vmulps %xmm1, %xmm0, %xmm2
620; HASWELL-NO-FMA-NEXT:    vbroadcastss {{.*#+}} xmm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
621; HASWELL-NO-FMA-NEXT:    vsubps %xmm2, %xmm3, %xmm2
622; HASWELL-NO-FMA-NEXT:    vmulps %xmm2, %xmm1, %xmm2
623; HASWELL-NO-FMA-NEXT:    vaddps %xmm2, %xmm1, %xmm1
624; HASWELL-NO-FMA-NEXT:    vmulps %xmm1, %xmm0, %xmm0
625; HASWELL-NO-FMA-NEXT:    vsubps %xmm0, %xmm3, %xmm0
626; HASWELL-NO-FMA-NEXT:    vmulps %xmm0, %xmm1, %xmm0
627; HASWELL-NO-FMA-NEXT:    vaddps %xmm0, %xmm1, %xmm0
628; HASWELL-NO-FMA-NEXT:    retq
629;
630; AVX512-LABEL: v4f32_two_step:
631; AVX512:       # %bb.0:
632; AVX512-NEXT:    vrcpps %xmm0, %xmm1
633; AVX512-NEXT:    vbroadcastss {{.*#+}} xmm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
634; AVX512-NEXT:    vmovaps %xmm1, %xmm3
635; AVX512-NEXT:    vfmsub213ps {{.*#+}} xmm3 = (xmm0 * xmm3) - xmm2
636; AVX512-NEXT:    vfnmadd132ps {{.*#+}} xmm3 = -(xmm3 * xmm1) + xmm1
637; AVX512-NEXT:    vfmsub213ps {{.*#+}} xmm0 = (xmm3 * xmm0) - xmm2
638; AVX512-NEXT:    vfnmadd132ps {{.*#+}} xmm0 = -(xmm0 * xmm3) + xmm3
639; AVX512-NEXT:    retq
640  %div = fdiv fast <4 x float> <float 1.0, float 1.0, float 1.0, float 1.0>, %x
641  ret <4 x float> %div
642}
643
644define <8 x float> @v8f32_no_estimate(<8 x float> %x) #0 {
645; SSE-LABEL: v8f32_no_estimate:
646; SSE:       # %bb.0:
647; SSE-NEXT:    movaps {{.*#+}} xmm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
648; SSE-NEXT:    movaps %xmm2, %xmm3
649; SSE-NEXT:    divps %xmm0, %xmm3
650; SSE-NEXT:    divps %xmm1, %xmm2
651; SSE-NEXT:    movaps %xmm3, %xmm0
652; SSE-NEXT:    movaps %xmm2, %xmm1
653; SSE-NEXT:    retq
654;
655; AVX-LABEL: v8f32_no_estimate:
656; AVX:       # %bb.0:
657; AVX-NEXT:    vbroadcastss {{.*#+}} ymm1 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
658; AVX-NEXT:    vdivps %ymm0, %ymm1, %ymm0
659; AVX-NEXT:    retq
660  %div = fdiv fast <8 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, %x
661  ret <8 x float> %div
662}
663
664define <8 x float> @v8f32_one_step(<8 x float> %x) #1 {
665; SSE-LABEL: v8f32_one_step:
666; SSE:       # %bb.0:
667; SSE-NEXT:    rcpps %xmm0, %xmm4
668; SSE-NEXT:    mulps %xmm4, %xmm0
669; SSE-NEXT:    movaps {{.*#+}} xmm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
670; SSE-NEXT:    movaps %xmm2, %xmm3
671; SSE-NEXT:    subps %xmm0, %xmm3
672; SSE-NEXT:    mulps %xmm4, %xmm3
673; SSE-NEXT:    addps %xmm4, %xmm3
674; SSE-NEXT:    rcpps %xmm1, %xmm0
675; SSE-NEXT:    mulps %xmm0, %xmm1
676; SSE-NEXT:    subps %xmm1, %xmm2
677; SSE-NEXT:    mulps %xmm0, %xmm2
678; SSE-NEXT:    addps %xmm0, %xmm2
679; SSE-NEXT:    movaps %xmm3, %xmm0
680; SSE-NEXT:    movaps %xmm2, %xmm1
681; SSE-NEXT:    retq
682;
683; AVX-RECIP-LABEL: v8f32_one_step:
684; AVX-RECIP:       # %bb.0:
685; AVX-RECIP-NEXT:    vrcpps %ymm0, %ymm1
686; AVX-RECIP-NEXT:    vmulps %ymm1, %ymm0, %ymm0
687; AVX-RECIP-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
688; AVX-RECIP-NEXT:    vsubps %ymm0, %ymm2, %ymm0
689; AVX-RECIP-NEXT:    vmulps %ymm0, %ymm1, %ymm0
690; AVX-RECIP-NEXT:    vaddps %ymm0, %ymm1, %ymm0
691; AVX-RECIP-NEXT:    retq
692;
693; FMA-RECIP-LABEL: v8f32_one_step:
694; FMA-RECIP:       # %bb.0:
695; FMA-RECIP-NEXT:    vrcpps %ymm0, %ymm1
696; FMA-RECIP-NEXT:    vfmsub213ps {{.*#+}} ymm0 = (ymm1 * ymm0) - mem
697; FMA-RECIP-NEXT:    vfnmadd132ps {{.*#+}} ymm0 = -(ymm0 * ymm1) + ymm1
698; FMA-RECIP-NEXT:    retq
699;
700; BDVER2-LABEL: v8f32_one_step:
701; BDVER2:       # %bb.0:
702; BDVER2-NEXT:    vrcpps %ymm0, %ymm1
703; BDVER2-NEXT:    vfmsubps {{.*#+}} ymm0 = (ymm0 * ymm1) - mem
704; BDVER2-NEXT:    vfnmaddps {{.*#+}} ymm0 = -(ymm1 * ymm0) + ymm1
705; BDVER2-NEXT:    retq
706;
707; BTVER2-LABEL: v8f32_one_step:
708; BTVER2:       # %bb.0:
709; BTVER2-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
710; BTVER2-NEXT:    vrcpps %ymm0, %ymm1
711; BTVER2-NEXT:    vmulps %ymm1, %ymm0, %ymm0
712; BTVER2-NEXT:    vsubps %ymm0, %ymm2, %ymm0
713; BTVER2-NEXT:    vmulps %ymm0, %ymm1, %ymm0
714; BTVER2-NEXT:    vaddps %ymm0, %ymm1, %ymm0
715; BTVER2-NEXT:    retq
716;
717; SANDY-LABEL: v8f32_one_step:
718; SANDY:       # %bb.0:
719; SANDY-NEXT:    vrcpps %ymm0, %ymm1
720; SANDY-NEXT:    vmulps %ymm1, %ymm0, %ymm0
721; SANDY-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
722; SANDY-NEXT:    vsubps %ymm0, %ymm2, %ymm0
723; SANDY-NEXT:    vmulps %ymm0, %ymm1, %ymm0
724; SANDY-NEXT:    vaddps %ymm0, %ymm1, %ymm0
725; SANDY-NEXT:    retq
726;
727; HASWELL-LABEL: v8f32_one_step:
728; HASWELL:       # %bb.0:
729; HASWELL-NEXT:    vrcpps %ymm0, %ymm1
730; HASWELL-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
731; HASWELL-NEXT:    vfmsub213ps {{.*#+}} ymm0 = (ymm1 * ymm0) - ymm2
732; HASWELL-NEXT:    vfnmadd132ps {{.*#+}} ymm0 = -(ymm0 * ymm1) + ymm1
733; HASWELL-NEXT:    retq
734;
735; HASWELL-NO-FMA-LABEL: v8f32_one_step:
736; HASWELL-NO-FMA:       # %bb.0:
737; HASWELL-NO-FMA-NEXT:    vrcpps %ymm0, %ymm1
738; HASWELL-NO-FMA-NEXT:    vmulps %ymm1, %ymm0, %ymm0
739; HASWELL-NO-FMA-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
740; HASWELL-NO-FMA-NEXT:    vsubps %ymm0, %ymm2, %ymm0
741; HASWELL-NO-FMA-NEXT:    vmulps %ymm0, %ymm1, %ymm0
742; HASWELL-NO-FMA-NEXT:    vaddps %ymm0, %ymm1, %ymm0
743; HASWELL-NO-FMA-NEXT:    retq
744;
745; KNL-LABEL: v8f32_one_step:
746; KNL:       # %bb.0:
747; KNL-NEXT:    vrcpps %ymm0, %ymm1
748; KNL-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
749; KNL-NEXT:    vfmsub213ps {{.*#+}} ymm0 = (ymm1 * ymm0) - ymm2
750; KNL-NEXT:    vfnmadd132ps {{.*#+}} ymm0 = -(ymm0 * ymm1) + ymm1
751; KNL-NEXT:    retq
752;
753; SKX-LABEL: v8f32_one_step:
754; SKX:       # %bb.0:
755; SKX-NEXT:    vrcpps %ymm0, %ymm1
756; SKX-NEXT:    vfmsub213ps {{.*#+}} ymm0 = (ymm1 * ymm0) - mem
757; SKX-NEXT:    vfnmadd132ps {{.*#+}} ymm0 = -(ymm0 * ymm1) + ymm1
758; SKX-NEXT:    retq
759  %div = fdiv fast <8 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, %x
760  ret <8 x float> %div
761}
762
763define <8 x float> @v8f32_two_step(<8 x float> %x) #2 {
764; SSE-LABEL: v8f32_two_step:
765; SSE:       # %bb.0:
766; SSE-NEXT:    movaps %xmm1, %xmm2
767; SSE-NEXT:    rcpps %xmm0, %xmm3
768; SSE-NEXT:    movaps %xmm0, %xmm4
769; SSE-NEXT:    mulps %xmm3, %xmm4
770; SSE-NEXT:    movaps {{.*#+}} xmm1 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
771; SSE-NEXT:    movaps %xmm1, %xmm5
772; SSE-NEXT:    subps %xmm4, %xmm5
773; SSE-NEXT:    mulps %xmm3, %xmm5
774; SSE-NEXT:    addps %xmm3, %xmm5
775; SSE-NEXT:    mulps %xmm5, %xmm0
776; SSE-NEXT:    movaps %xmm1, %xmm3
777; SSE-NEXT:    subps %xmm0, %xmm3
778; SSE-NEXT:    mulps %xmm5, %xmm3
779; SSE-NEXT:    addps %xmm5, %xmm3
780; SSE-NEXT:    rcpps %xmm2, %xmm0
781; SSE-NEXT:    movaps %xmm2, %xmm4
782; SSE-NEXT:    mulps %xmm0, %xmm4
783; SSE-NEXT:    movaps %xmm1, %xmm5
784; SSE-NEXT:    subps %xmm4, %xmm5
785; SSE-NEXT:    mulps %xmm0, %xmm5
786; SSE-NEXT:    addps %xmm0, %xmm5
787; SSE-NEXT:    mulps %xmm5, %xmm2
788; SSE-NEXT:    subps %xmm2, %xmm1
789; SSE-NEXT:    mulps %xmm5, %xmm1
790; SSE-NEXT:    addps %xmm5, %xmm1
791; SSE-NEXT:    movaps %xmm3, %xmm0
792; SSE-NEXT:    retq
793;
794; AVX-RECIP-LABEL: v8f32_two_step:
795; AVX-RECIP:       # %bb.0:
796; AVX-RECIP-NEXT:    vrcpps %ymm0, %ymm1
797; AVX-RECIP-NEXT:    vmulps %ymm1, %ymm0, %ymm2
798; AVX-RECIP-NEXT:    vbroadcastss {{.*#+}} ymm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
799; AVX-RECIP-NEXT:    vsubps %ymm2, %ymm3, %ymm2
800; AVX-RECIP-NEXT:    vmulps %ymm2, %ymm1, %ymm2
801; AVX-RECIP-NEXT:    vaddps %ymm2, %ymm1, %ymm1
802; AVX-RECIP-NEXT:    vmulps %ymm1, %ymm0, %ymm0
803; AVX-RECIP-NEXT:    vsubps %ymm0, %ymm3, %ymm0
804; AVX-RECIP-NEXT:    vmulps %ymm0, %ymm1, %ymm0
805; AVX-RECIP-NEXT:    vaddps %ymm0, %ymm1, %ymm0
806; AVX-RECIP-NEXT:    retq
807;
808; FMA-RECIP-LABEL: v8f32_two_step:
809; FMA-RECIP:       # %bb.0:
810; FMA-RECIP-NEXT:    vrcpps %ymm0, %ymm1
811; FMA-RECIP-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
812; FMA-RECIP-NEXT:    vmovaps %ymm1, %ymm3
813; FMA-RECIP-NEXT:    vfmsub213ps {{.*#+}} ymm3 = (ymm0 * ymm3) - ymm2
814; FMA-RECIP-NEXT:    vfnmadd132ps {{.*#+}} ymm3 = -(ymm3 * ymm1) + ymm1
815; FMA-RECIP-NEXT:    vfmsub213ps {{.*#+}} ymm0 = (ymm3 * ymm0) - ymm2
816; FMA-RECIP-NEXT:    vfnmadd132ps {{.*#+}} ymm0 = -(ymm0 * ymm3) + ymm3
817; FMA-RECIP-NEXT:    retq
818;
819; BDVER2-LABEL: v8f32_two_step:
820; BDVER2:       # %bb.0:
821; BDVER2-NEXT:    vrcpps %ymm0, %ymm1
822; BDVER2-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
823; BDVER2-NEXT:    vfmsubps {{.*#+}} ymm3 = (ymm0 * ymm1) - ymm2
824; BDVER2-NEXT:    vfnmaddps {{.*#+}} ymm1 = -(ymm1 * ymm3) + ymm1
825; BDVER2-NEXT:    vfmsubps {{.*#+}} ymm0 = (ymm0 * ymm1) - ymm2
826; BDVER2-NEXT:    vfnmaddps {{.*#+}} ymm0 = -(ymm1 * ymm0) + ymm1
827; BDVER2-NEXT:    retq
828;
829; BTVER2-LABEL: v8f32_two_step:
830; BTVER2:       # %bb.0:
831; BTVER2-NEXT:    vbroadcastss {{.*#+}} ymm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
832; BTVER2-NEXT:    vrcpps %ymm0, %ymm1
833; BTVER2-NEXT:    vmulps %ymm1, %ymm0, %ymm2
834; BTVER2-NEXT:    vsubps %ymm2, %ymm3, %ymm2
835; BTVER2-NEXT:    vmulps %ymm2, %ymm1, %ymm2
836; BTVER2-NEXT:    vaddps %ymm2, %ymm1, %ymm1
837; BTVER2-NEXT:    vmulps %ymm1, %ymm0, %ymm0
838; BTVER2-NEXT:    vsubps %ymm0, %ymm3, %ymm0
839; BTVER2-NEXT:    vmulps %ymm0, %ymm1, %ymm0
840; BTVER2-NEXT:    vaddps %ymm0, %ymm1, %ymm0
841; BTVER2-NEXT:    retq
842;
843; SANDY-LABEL: v8f32_two_step:
844; SANDY:       # %bb.0:
845; SANDY-NEXT:    vrcpps %ymm0, %ymm1
846; SANDY-NEXT:    vmulps %ymm1, %ymm0, %ymm2
847; SANDY-NEXT:    vbroadcastss {{.*#+}} ymm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
848; SANDY-NEXT:    vsubps %ymm2, %ymm3, %ymm2
849; SANDY-NEXT:    vmulps %ymm2, %ymm1, %ymm2
850; SANDY-NEXT:    vaddps %ymm2, %ymm1, %ymm1
851; SANDY-NEXT:    vmulps %ymm1, %ymm0, %ymm0
852; SANDY-NEXT:    vsubps %ymm0, %ymm3, %ymm0
853; SANDY-NEXT:    vmulps %ymm0, %ymm1, %ymm0
854; SANDY-NEXT:    vaddps %ymm0, %ymm1, %ymm0
855; SANDY-NEXT:    retq
856;
857; HASWELL-LABEL: v8f32_two_step:
858; HASWELL:       # %bb.0:
859; HASWELL-NEXT:    vrcpps %ymm0, %ymm1
860; HASWELL-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
861; HASWELL-NEXT:    vmovaps %ymm1, %ymm3
862; HASWELL-NEXT:    vfmsub213ps {{.*#+}} ymm3 = (ymm0 * ymm3) - ymm2
863; HASWELL-NEXT:    vfnmadd132ps {{.*#+}} ymm3 = -(ymm3 * ymm1) + ymm1
864; HASWELL-NEXT:    vfmsub213ps {{.*#+}} ymm0 = (ymm3 * ymm0) - ymm2
865; HASWELL-NEXT:    vfnmadd132ps {{.*#+}} ymm0 = -(ymm0 * ymm3) + ymm3
866; HASWELL-NEXT:    retq
867;
868; HASWELL-NO-FMA-LABEL: v8f32_two_step:
869; HASWELL-NO-FMA:       # %bb.0:
870; HASWELL-NO-FMA-NEXT:    vrcpps %ymm0, %ymm1
871; HASWELL-NO-FMA-NEXT:    vmulps %ymm1, %ymm0, %ymm2
872; HASWELL-NO-FMA-NEXT:    vbroadcastss {{.*#+}} ymm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
873; HASWELL-NO-FMA-NEXT:    vsubps %ymm2, %ymm3, %ymm2
874; HASWELL-NO-FMA-NEXT:    vmulps %ymm2, %ymm1, %ymm2
875; HASWELL-NO-FMA-NEXT:    vaddps %ymm2, %ymm1, %ymm1
876; HASWELL-NO-FMA-NEXT:    vmulps %ymm1, %ymm0, %ymm0
877; HASWELL-NO-FMA-NEXT:    vsubps %ymm0, %ymm3, %ymm0
878; HASWELL-NO-FMA-NEXT:    vmulps %ymm0, %ymm1, %ymm0
879; HASWELL-NO-FMA-NEXT:    vaddps %ymm0, %ymm1, %ymm0
880; HASWELL-NO-FMA-NEXT:    retq
881;
882; AVX512-LABEL: v8f32_two_step:
883; AVX512:       # %bb.0:
884; AVX512-NEXT:    vrcpps %ymm0, %ymm1
885; AVX512-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
886; AVX512-NEXT:    vmovaps %ymm1, %ymm3
887; AVX512-NEXT:    vfmsub213ps {{.*#+}} ymm3 = (ymm0 * ymm3) - ymm2
888; AVX512-NEXT:    vfnmadd132ps {{.*#+}} ymm3 = -(ymm3 * ymm1) + ymm1
889; AVX512-NEXT:    vfmsub213ps {{.*#+}} ymm0 = (ymm3 * ymm0) - ymm2
890; AVX512-NEXT:    vfnmadd132ps {{.*#+}} ymm0 = -(ymm0 * ymm3) + ymm3
891; AVX512-NEXT:    retq
892  %div = fdiv fast <8 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, %x
893  ret <8 x float> %div
894}
895
896define <16 x float> @v16f32_no_estimate(<16 x float> %x) #0 {
897; SSE-LABEL: v16f32_no_estimate:
898; SSE:       # %bb.0:
899; SSE-NEXT:    movaps {{.*#+}} xmm4 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
900; SSE-NEXT:    movaps %xmm4, %xmm5
901; SSE-NEXT:    divps %xmm0, %xmm5
902; SSE-NEXT:    movaps %xmm4, %xmm6
903; SSE-NEXT:    divps %xmm1, %xmm6
904; SSE-NEXT:    movaps %xmm4, %xmm7
905; SSE-NEXT:    divps %xmm2, %xmm7
906; SSE-NEXT:    divps %xmm3, %xmm4
907; SSE-NEXT:    movaps %xmm5, %xmm0
908; SSE-NEXT:    movaps %xmm6, %xmm1
909; SSE-NEXT:    movaps %xmm7, %xmm2
910; SSE-NEXT:    movaps %xmm4, %xmm3
911; SSE-NEXT:    retq
912;
913; AVX-RECIP-LABEL: v16f32_no_estimate:
914; AVX-RECIP:       # %bb.0:
915; AVX-RECIP-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
916; AVX-RECIP-NEXT:    vdivps %ymm0, %ymm2, %ymm0
917; AVX-RECIP-NEXT:    vdivps %ymm1, %ymm2, %ymm1
918; AVX-RECIP-NEXT:    retq
919;
920; FMA-RECIP-LABEL: v16f32_no_estimate:
921; FMA-RECIP:       # %bb.0:
922; FMA-RECIP-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
923; FMA-RECIP-NEXT:    vdivps %ymm0, %ymm2, %ymm0
924; FMA-RECIP-NEXT:    vdivps %ymm1, %ymm2, %ymm1
925; FMA-RECIP-NEXT:    retq
926;
927; BDVER2-LABEL: v16f32_no_estimate:
928; BDVER2:       # %bb.0:
929; BDVER2-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
930; BDVER2-NEXT:    vdivps %ymm0, %ymm2, %ymm0
931; BDVER2-NEXT:    vdivps %ymm1, %ymm2, %ymm1
932; BDVER2-NEXT:    retq
933;
934; BTVER2-LABEL: v16f32_no_estimate:
935; BTVER2:       # %bb.0:
936; BTVER2-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
937; BTVER2-NEXT:    vdivps %ymm0, %ymm2, %ymm0
938; BTVER2-NEXT:    vdivps %ymm1, %ymm2, %ymm1
939; BTVER2-NEXT:    retq
940;
941; SANDY-LABEL: v16f32_no_estimate:
942; SANDY:       # %bb.0:
943; SANDY-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
944; SANDY-NEXT:    vdivps %ymm0, %ymm2, %ymm0
945; SANDY-NEXT:    vdivps %ymm1, %ymm2, %ymm1
946; SANDY-NEXT:    retq
947;
948; HASWELL-LABEL: v16f32_no_estimate:
949; HASWELL:       # %bb.0:
950; HASWELL-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
951; HASWELL-NEXT:    vdivps %ymm0, %ymm2, %ymm0
952; HASWELL-NEXT:    vdivps %ymm1, %ymm2, %ymm1
953; HASWELL-NEXT:    retq
954;
955; HASWELL-NO-FMA-LABEL: v16f32_no_estimate:
956; HASWELL-NO-FMA:       # %bb.0:
957; HASWELL-NO-FMA-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
958; HASWELL-NO-FMA-NEXT:    vdivps %ymm0, %ymm2, %ymm0
959; HASWELL-NO-FMA-NEXT:    vdivps %ymm1, %ymm2, %ymm1
960; HASWELL-NO-FMA-NEXT:    retq
961;
962; AVX512-LABEL: v16f32_no_estimate:
963; AVX512:       # %bb.0:
964; AVX512-NEXT:    vbroadcastss {{.*#+}} zmm1 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
965; AVX512-NEXT:    vdivps %zmm0, %zmm1, %zmm0
966; AVX512-NEXT:    retq
967  %div = fdiv fast <16 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, %x
968  ret <16 x float> %div
969}
970
971define <16 x float> @v16f32_one_step(<16 x float> %x) #1 {
972; SSE-LABEL: v16f32_one_step:
973; SSE:       # %bb.0:
974; SSE-NEXT:    movaps %xmm3, %xmm4
975; SSE-NEXT:    movaps %xmm0, %xmm5
976; SSE-NEXT:    rcpps %xmm0, %xmm6
977; SSE-NEXT:    mulps %xmm6, %xmm5
978; SSE-NEXT:    movaps {{.*#+}} xmm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
979; SSE-NEXT:    movaps %xmm3, %xmm0
980; SSE-NEXT:    subps %xmm5, %xmm0
981; SSE-NEXT:    mulps %xmm6, %xmm0
982; SSE-NEXT:    addps %xmm6, %xmm0
983; SSE-NEXT:    rcpps %xmm1, %xmm6
984; SSE-NEXT:    mulps %xmm6, %xmm1
985; SSE-NEXT:    movaps %xmm3, %xmm5
986; SSE-NEXT:    subps %xmm1, %xmm5
987; SSE-NEXT:    mulps %xmm6, %xmm5
988; SSE-NEXT:    addps %xmm6, %xmm5
989; SSE-NEXT:    rcpps %xmm2, %xmm1
990; SSE-NEXT:    mulps %xmm1, %xmm2
991; SSE-NEXT:    movaps %xmm3, %xmm6
992; SSE-NEXT:    subps %xmm2, %xmm6
993; SSE-NEXT:    mulps %xmm1, %xmm6
994; SSE-NEXT:    addps %xmm1, %xmm6
995; SSE-NEXT:    rcpps %xmm4, %xmm1
996; SSE-NEXT:    mulps %xmm1, %xmm4
997; SSE-NEXT:    subps %xmm4, %xmm3
998; SSE-NEXT:    mulps %xmm1, %xmm3
999; SSE-NEXT:    addps %xmm1, %xmm3
1000; SSE-NEXT:    movaps %xmm5, %xmm1
1001; SSE-NEXT:    movaps %xmm6, %xmm2
1002; SSE-NEXT:    retq
1003;
1004; AVX-RECIP-LABEL: v16f32_one_step:
1005; AVX-RECIP:       # %bb.0:
1006; AVX-RECIP-NEXT:    vrcpps %ymm0, %ymm2
1007; AVX-RECIP-NEXT:    vmulps %ymm2, %ymm0, %ymm0
1008; AVX-RECIP-NEXT:    vbroadcastss {{.*#+}} ymm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
1009; AVX-RECIP-NEXT:    vsubps %ymm0, %ymm3, %ymm0
1010; AVX-RECIP-NEXT:    vmulps %ymm0, %ymm2, %ymm0
1011; AVX-RECIP-NEXT:    vaddps %ymm0, %ymm2, %ymm0
1012; AVX-RECIP-NEXT:    vrcpps %ymm1, %ymm2
1013; AVX-RECIP-NEXT:    vmulps %ymm2, %ymm1, %ymm1
1014; AVX-RECIP-NEXT:    vsubps %ymm1, %ymm3, %ymm1
1015; AVX-RECIP-NEXT:    vmulps %ymm1, %ymm2, %ymm1
1016; AVX-RECIP-NEXT:    vaddps %ymm1, %ymm2, %ymm1
1017; AVX-RECIP-NEXT:    retq
1018;
1019; FMA-RECIP-LABEL: v16f32_one_step:
1020; FMA-RECIP:       # %bb.0:
1021; FMA-RECIP-NEXT:    vrcpps %ymm0, %ymm2
1022; FMA-RECIP-NEXT:    vbroadcastss {{.*#+}} ymm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
1023; FMA-RECIP-NEXT:    vfmsub213ps {{.*#+}} ymm0 = (ymm2 * ymm0) - ymm3
1024; FMA-RECIP-NEXT:    vfnmadd132ps {{.*#+}} ymm0 = -(ymm0 * ymm2) + ymm2
1025; FMA-RECIP-NEXT:    vrcpps %ymm1, %ymm2
1026; FMA-RECIP-NEXT:    vfmsub213ps {{.*#+}} ymm1 = (ymm2 * ymm1) - ymm3
1027; FMA-RECIP-NEXT:    vfnmadd132ps {{.*#+}} ymm1 = -(ymm1 * ymm2) + ymm2
1028; FMA-RECIP-NEXT:    retq
1029;
1030; BDVER2-LABEL: v16f32_one_step:
1031; BDVER2:       # %bb.0:
1032; BDVER2-NEXT:    vrcpps %ymm0, %ymm2
1033; BDVER2-NEXT:    vbroadcastss {{.*#+}} ymm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
1034; BDVER2-NEXT:    vrcpps %ymm1, %ymm4
1035; BDVER2-NEXT:    vfmsubps {{.*#+}} ymm0 = (ymm0 * ymm2) - ymm3
1036; BDVER2-NEXT:    vfmsubps {{.*#+}} ymm1 = (ymm1 * ymm4) - ymm3
1037; BDVER2-NEXT:    vfnmaddps {{.*#+}} ymm0 = -(ymm2 * ymm0) + ymm2
1038; BDVER2-NEXT:    vfnmaddps {{.*#+}} ymm1 = -(ymm4 * ymm1) + ymm4
1039; BDVER2-NEXT:    retq
1040;
1041; BTVER2-LABEL: v16f32_one_step:
1042; BTVER2:       # %bb.0:
1043; BTVER2-NEXT:    vbroadcastss {{.*#+}} ymm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
1044; BTVER2-NEXT:    vrcpps %ymm0, %ymm2
1045; BTVER2-NEXT:    vrcpps %ymm1, %ymm4
1046; BTVER2-NEXT:    vmulps %ymm2, %ymm0, %ymm0
1047; BTVER2-NEXT:    vmulps %ymm4, %ymm1, %ymm1
1048; BTVER2-NEXT:    vsubps %ymm0, %ymm3, %ymm0
1049; BTVER2-NEXT:    vsubps %ymm1, %ymm3, %ymm1
1050; BTVER2-NEXT:    vmulps %ymm0, %ymm2, %ymm0
1051; BTVER2-NEXT:    vmulps %ymm1, %ymm4, %ymm1
1052; BTVER2-NEXT:    vaddps %ymm0, %ymm2, %ymm0
1053; BTVER2-NEXT:    vaddps %ymm1, %ymm4, %ymm1
1054; BTVER2-NEXT:    retq
1055;
1056; SANDY-LABEL: v16f32_one_step:
1057; SANDY:       # %bb.0:
1058; SANDY-NEXT:    vrcpps %ymm0, %ymm2
1059; SANDY-NEXT:    vmulps %ymm2, %ymm0, %ymm0
1060; SANDY-NEXT:    vbroadcastss {{.*#+}} ymm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
1061; SANDY-NEXT:    vsubps %ymm0, %ymm3, %ymm0
1062; SANDY-NEXT:    vrcpps %ymm1, %ymm4
1063; SANDY-NEXT:    vmulps %ymm0, %ymm2, %ymm0
1064; SANDY-NEXT:    vaddps %ymm0, %ymm2, %ymm0
1065; SANDY-NEXT:    vmulps %ymm4, %ymm1, %ymm1
1066; SANDY-NEXT:    vsubps %ymm1, %ymm3, %ymm1
1067; SANDY-NEXT:    vmulps %ymm1, %ymm4, %ymm1
1068; SANDY-NEXT:    vaddps %ymm1, %ymm4, %ymm1
1069; SANDY-NEXT:    retq
1070;
1071; HASWELL-LABEL: v16f32_one_step:
1072; HASWELL:       # %bb.0:
1073; HASWELL-NEXT:    vrcpps %ymm0, %ymm2
1074; HASWELL-NEXT:    vbroadcastss {{.*#+}} ymm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
1075; HASWELL-NEXT:    vrcpps %ymm1, %ymm4
1076; HASWELL-NEXT:    vfmsub213ps {{.*#+}} ymm0 = (ymm2 * ymm0) - ymm3
1077; HASWELL-NEXT:    vfnmadd132ps {{.*#+}} ymm0 = -(ymm0 * ymm2) + ymm2
1078; HASWELL-NEXT:    vfmsub213ps {{.*#+}} ymm1 = (ymm4 * ymm1) - ymm3
1079; HASWELL-NEXT:    vfnmadd132ps {{.*#+}} ymm1 = -(ymm1 * ymm4) + ymm4
1080; HASWELL-NEXT:    retq
1081;
1082; HASWELL-NO-FMA-LABEL: v16f32_one_step:
1083; HASWELL-NO-FMA:       # %bb.0:
1084; HASWELL-NO-FMA-NEXT:    vrcpps %ymm0, %ymm2
1085; HASWELL-NO-FMA-NEXT:    vmulps %ymm2, %ymm0, %ymm0
1086; HASWELL-NO-FMA-NEXT:    vbroadcastss {{.*#+}} ymm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
1087; HASWELL-NO-FMA-NEXT:    vsubps %ymm0, %ymm3, %ymm0
1088; HASWELL-NO-FMA-NEXT:    vmulps %ymm0, %ymm2, %ymm0
1089; HASWELL-NO-FMA-NEXT:    vaddps %ymm0, %ymm2, %ymm0
1090; HASWELL-NO-FMA-NEXT:    vrcpps %ymm1, %ymm2
1091; HASWELL-NO-FMA-NEXT:    vmulps %ymm2, %ymm1, %ymm1
1092; HASWELL-NO-FMA-NEXT:    vsubps %ymm1, %ymm3, %ymm1
1093; HASWELL-NO-FMA-NEXT:    vmulps %ymm1, %ymm2, %ymm1
1094; HASWELL-NO-FMA-NEXT:    vaddps %ymm1, %ymm2, %ymm1
1095; HASWELL-NO-FMA-NEXT:    retq
1096;
1097; AVX512-LABEL: v16f32_one_step:
1098; AVX512:       # %bb.0:
1099; AVX512-NEXT:    vrcp14ps %zmm0, %zmm1
1100; AVX512-NEXT:    vfmsub213ps {{.*#+}} zmm0 = (zmm1 * zmm0) - mem
1101; AVX512-NEXT:    vfnmadd132ps {{.*#+}} zmm0 = -(zmm0 * zmm1) + zmm1
1102; AVX512-NEXT:    retq
1103  %div = fdiv fast <16 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, %x
1104  ret <16 x float> %div
1105}
1106
1107define <16 x float> @v16f32_two_step(<16 x float> %x) #2 {
1108; SSE-LABEL: v16f32_two_step:
1109; SSE:       # %bb.0:
1110; SSE-NEXT:    movaps %xmm3, %xmm4
1111; SSE-NEXT:    movaps %xmm1, %xmm5
1112; SSE-NEXT:    movaps %xmm0, %xmm1
1113; SSE-NEXT:    rcpps %xmm0, %xmm0
1114; SSE-NEXT:    movaps %xmm1, %xmm6
1115; SSE-NEXT:    mulps %xmm0, %xmm6
1116; SSE-NEXT:    movaps {{.*#+}} xmm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
1117; SSE-NEXT:    movaps %xmm3, %xmm7
1118; SSE-NEXT:    subps %xmm6, %xmm7
1119; SSE-NEXT:    mulps %xmm0, %xmm7
1120; SSE-NEXT:    addps %xmm0, %xmm7
1121; SSE-NEXT:    mulps %xmm7, %xmm1
1122; SSE-NEXT:    movaps %xmm3, %xmm0
1123; SSE-NEXT:    subps %xmm1, %xmm0
1124; SSE-NEXT:    mulps %xmm7, %xmm0
1125; SSE-NEXT:    addps %xmm7, %xmm0
1126; SSE-NEXT:    rcpps %xmm5, %xmm1
1127; SSE-NEXT:    movaps %xmm5, %xmm6
1128; SSE-NEXT:    mulps %xmm1, %xmm6
1129; SSE-NEXT:    movaps %xmm3, %xmm7
1130; SSE-NEXT:    subps %xmm6, %xmm7
1131; SSE-NEXT:    mulps %xmm1, %xmm7
1132; SSE-NEXT:    addps %xmm1, %xmm7
1133; SSE-NEXT:    mulps %xmm7, %xmm5
1134; SSE-NEXT:    movaps %xmm3, %xmm1
1135; SSE-NEXT:    subps %xmm5, %xmm1
1136; SSE-NEXT:    mulps %xmm7, %xmm1
1137; SSE-NEXT:    addps %xmm7, %xmm1
1138; SSE-NEXT:    rcpps %xmm2, %xmm5
1139; SSE-NEXT:    movaps %xmm2, %xmm6
1140; SSE-NEXT:    mulps %xmm5, %xmm6
1141; SSE-NEXT:    movaps %xmm3, %xmm7
1142; SSE-NEXT:    subps %xmm6, %xmm7
1143; SSE-NEXT:    mulps %xmm5, %xmm7
1144; SSE-NEXT:    addps %xmm5, %xmm7
1145; SSE-NEXT:    mulps %xmm7, %xmm2
1146; SSE-NEXT:    movaps %xmm3, %xmm5
1147; SSE-NEXT:    subps %xmm2, %xmm5
1148; SSE-NEXT:    mulps %xmm7, %xmm5
1149; SSE-NEXT:    addps %xmm7, %xmm5
1150; SSE-NEXT:    rcpps %xmm4, %xmm2
1151; SSE-NEXT:    movaps %xmm4, %xmm6
1152; SSE-NEXT:    mulps %xmm2, %xmm6
1153; SSE-NEXT:    movaps %xmm3, %xmm7
1154; SSE-NEXT:    subps %xmm6, %xmm7
1155; SSE-NEXT:    mulps %xmm2, %xmm7
1156; SSE-NEXT:    addps %xmm2, %xmm7
1157; SSE-NEXT:    mulps %xmm7, %xmm4
1158; SSE-NEXT:    subps %xmm4, %xmm3
1159; SSE-NEXT:    mulps %xmm7, %xmm3
1160; SSE-NEXT:    addps %xmm7, %xmm3
1161; SSE-NEXT:    movaps %xmm5, %xmm2
1162; SSE-NEXT:    retq
1163;
1164; AVX-RECIP-LABEL: v16f32_two_step:
1165; AVX-RECIP:       # %bb.0:
1166; AVX-RECIP-NEXT:    vrcpps %ymm0, %ymm2
1167; AVX-RECIP-NEXT:    vmulps %ymm2, %ymm0, %ymm3
1168; AVX-RECIP-NEXT:    vbroadcastss {{.*#+}} ymm4 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
1169; AVX-RECIP-NEXT:    vsubps %ymm3, %ymm4, %ymm3
1170; AVX-RECIP-NEXT:    vmulps %ymm3, %ymm2, %ymm3
1171; AVX-RECIP-NEXT:    vaddps %ymm3, %ymm2, %ymm2
1172; AVX-RECIP-NEXT:    vmulps %ymm2, %ymm0, %ymm0
1173; AVX-RECIP-NEXT:    vsubps %ymm0, %ymm4, %ymm0
1174; AVX-RECIP-NEXT:    vmulps %ymm0, %ymm2, %ymm0
1175; AVX-RECIP-NEXT:    vaddps %ymm0, %ymm2, %ymm0
1176; AVX-RECIP-NEXT:    vrcpps %ymm1, %ymm2
1177; AVX-RECIP-NEXT:    vmulps %ymm2, %ymm1, %ymm3
1178; AVX-RECIP-NEXT:    vsubps %ymm3, %ymm4, %ymm3
1179; AVX-RECIP-NEXT:    vmulps %ymm3, %ymm2, %ymm3
1180; AVX-RECIP-NEXT:    vaddps %ymm3, %ymm2, %ymm2
1181; AVX-RECIP-NEXT:    vmulps %ymm2, %ymm1, %ymm1
1182; AVX-RECIP-NEXT:    vsubps %ymm1, %ymm4, %ymm1
1183; AVX-RECIP-NEXT:    vmulps %ymm1, %ymm2, %ymm1
1184; AVX-RECIP-NEXT:    vaddps %ymm1, %ymm2, %ymm1
1185; AVX-RECIP-NEXT:    retq
1186;
1187; FMA-RECIP-LABEL: v16f32_two_step:
1188; FMA-RECIP:       # %bb.0:
1189; FMA-RECIP-NEXT:    vrcpps %ymm0, %ymm2
1190; FMA-RECIP-NEXT:    vbroadcastss {{.*#+}} ymm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
1191; FMA-RECIP-NEXT:    vmovaps %ymm2, %ymm4
1192; FMA-RECIP-NEXT:    vfmsub213ps {{.*#+}} ymm4 = (ymm0 * ymm4) - ymm3
1193; FMA-RECIP-NEXT:    vfnmadd132ps {{.*#+}} ymm4 = -(ymm4 * ymm2) + ymm2
1194; FMA-RECIP-NEXT:    vfmsub213ps {{.*#+}} ymm0 = (ymm4 * ymm0) - ymm3
1195; FMA-RECIP-NEXT:    vfnmadd132ps {{.*#+}} ymm0 = -(ymm0 * ymm4) + ymm4
1196; FMA-RECIP-NEXT:    vrcpps %ymm1, %ymm2
1197; FMA-RECIP-NEXT:    vmovaps %ymm2, %ymm4
1198; FMA-RECIP-NEXT:    vfmsub213ps {{.*#+}} ymm4 = (ymm1 * ymm4) - ymm3
1199; FMA-RECIP-NEXT:    vfnmadd132ps {{.*#+}} ymm4 = -(ymm4 * ymm2) + ymm2
1200; FMA-RECIP-NEXT:    vfmsub213ps {{.*#+}} ymm1 = (ymm4 * ymm1) - ymm3
1201; FMA-RECIP-NEXT:    vfnmadd132ps {{.*#+}} ymm1 = -(ymm1 * ymm4) + ymm4
1202; FMA-RECIP-NEXT:    retq
1203;
1204; BDVER2-LABEL: v16f32_two_step:
1205; BDVER2:       # %bb.0:
1206; BDVER2-NEXT:    vrcpps %ymm0, %ymm2
1207; BDVER2-NEXT:    vbroadcastss {{.*#+}} ymm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
1208; BDVER2-NEXT:    vfmsubps {{.*#+}} ymm4 = (ymm0 * ymm2) - ymm3
1209; BDVER2-NEXT:    vfnmaddps {{.*#+}} ymm2 = -(ymm2 * ymm4) + ymm2
1210; BDVER2-NEXT:    vfmsubps {{.*#+}} ymm0 = (ymm0 * ymm2) - ymm3
1211; BDVER2-NEXT:    vfnmaddps {{.*#+}} ymm0 = -(ymm2 * ymm0) + ymm2
1212; BDVER2-NEXT:    vrcpps %ymm1, %ymm2
1213; BDVER2-NEXT:    vfmsubps {{.*#+}} ymm4 = (ymm1 * ymm2) - ymm3
1214; BDVER2-NEXT:    vfnmaddps {{.*#+}} ymm2 = -(ymm2 * ymm4) + ymm2
1215; BDVER2-NEXT:    vfmsubps {{.*#+}} ymm1 = (ymm1 * ymm2) - ymm3
1216; BDVER2-NEXT:    vfnmaddps {{.*#+}} ymm1 = -(ymm2 * ymm1) + ymm2
1217; BDVER2-NEXT:    retq
1218;
1219; BTVER2-LABEL: v16f32_two_step:
1220; BTVER2:       # %bb.0:
1221; BTVER2-NEXT:    vbroadcastss {{.*#+}} ymm4 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
1222; BTVER2-NEXT:    vrcpps %ymm0, %ymm2
1223; BTVER2-NEXT:    vmulps %ymm2, %ymm0, %ymm3
1224; BTVER2-NEXT:    vsubps %ymm3, %ymm4, %ymm3
1225; BTVER2-NEXT:    vmulps %ymm3, %ymm2, %ymm3
1226; BTVER2-NEXT:    vaddps %ymm3, %ymm2, %ymm2
1227; BTVER2-NEXT:    vmulps %ymm2, %ymm0, %ymm0
1228; BTVER2-NEXT:    vsubps %ymm0, %ymm4, %ymm0
1229; BTVER2-NEXT:    vmulps %ymm0, %ymm2, %ymm0
1230; BTVER2-NEXT:    vaddps %ymm0, %ymm2, %ymm0
1231; BTVER2-NEXT:    vrcpps %ymm1, %ymm2
1232; BTVER2-NEXT:    vmulps %ymm2, %ymm1, %ymm3
1233; BTVER2-NEXT:    vsubps %ymm3, %ymm4, %ymm3
1234; BTVER2-NEXT:    vmulps %ymm3, %ymm2, %ymm3
1235; BTVER2-NEXT:    vaddps %ymm3, %ymm2, %ymm2
1236; BTVER2-NEXT:    vmulps %ymm2, %ymm1, %ymm1
1237; BTVER2-NEXT:    vsubps %ymm1, %ymm4, %ymm1
1238; BTVER2-NEXT:    vmulps %ymm1, %ymm2, %ymm1
1239; BTVER2-NEXT:    vaddps %ymm1, %ymm2, %ymm1
1240; BTVER2-NEXT:    retq
1241;
1242; SANDY-LABEL: v16f32_two_step:
1243; SANDY:       # %bb.0:
1244; SANDY-NEXT:    vrcpps %ymm0, %ymm2
1245; SANDY-NEXT:    vmulps %ymm2, %ymm0, %ymm3
1246; SANDY-NEXT:    vbroadcastss {{.*#+}} ymm4 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
1247; SANDY-NEXT:    vsubps %ymm3, %ymm4, %ymm3
1248; SANDY-NEXT:    vmulps %ymm3, %ymm2, %ymm3
1249; SANDY-NEXT:    vaddps %ymm3, %ymm2, %ymm2
1250; SANDY-NEXT:    vmulps %ymm2, %ymm0, %ymm0
1251; SANDY-NEXT:    vsubps %ymm0, %ymm4, %ymm0
1252; SANDY-NEXT:    vrcpps %ymm1, %ymm3
1253; SANDY-NEXT:    vmulps %ymm0, %ymm2, %ymm0
1254; SANDY-NEXT:    vaddps %ymm0, %ymm2, %ymm0
1255; SANDY-NEXT:    vmulps %ymm3, %ymm1, %ymm2
1256; SANDY-NEXT:    vsubps %ymm2, %ymm4, %ymm2
1257; SANDY-NEXT:    vmulps %ymm2, %ymm3, %ymm2
1258; SANDY-NEXT:    vaddps %ymm2, %ymm3, %ymm2
1259; SANDY-NEXT:    vmulps %ymm2, %ymm1, %ymm1
1260; SANDY-NEXT:    vsubps %ymm1, %ymm4, %ymm1
1261; SANDY-NEXT:    vmulps %ymm1, %ymm2, %ymm1
1262; SANDY-NEXT:    vaddps %ymm1, %ymm2, %ymm1
1263; SANDY-NEXT:    retq
1264;
1265; HASWELL-LABEL: v16f32_two_step:
1266; HASWELL:       # %bb.0:
1267; HASWELL-NEXT:    vrcpps %ymm0, %ymm2
1268; HASWELL-NEXT:    vbroadcastss {{.*#+}} ymm3 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
1269; HASWELL-NEXT:    vmovaps %ymm2, %ymm4
1270; HASWELL-NEXT:    vfmsub213ps {{.*#+}} ymm4 = (ymm0 * ymm4) - ymm3
1271; HASWELL-NEXT:    vfnmadd132ps {{.*#+}} ymm4 = -(ymm4 * ymm2) + ymm2
1272; HASWELL-NEXT:    vfmsub213ps {{.*#+}} ymm0 = (ymm4 * ymm0) - ymm3
1273; HASWELL-NEXT:    vfnmadd132ps {{.*#+}} ymm0 = -(ymm0 * ymm4) + ymm4
1274; HASWELL-NEXT:    vrcpps %ymm1, %ymm2
1275; HASWELL-NEXT:    vmovaps %ymm2, %ymm4
1276; HASWELL-NEXT:    vfmsub213ps {{.*#+}} ymm4 = (ymm1 * ymm4) - ymm3
1277; HASWELL-NEXT:    vfnmadd132ps {{.*#+}} ymm4 = -(ymm4 * ymm2) + ymm2
1278; HASWELL-NEXT:    vfmsub213ps {{.*#+}} ymm1 = (ymm4 * ymm1) - ymm3
1279; HASWELL-NEXT:    vfnmadd132ps {{.*#+}} ymm1 = -(ymm1 * ymm4) + ymm4
1280; HASWELL-NEXT:    retq
1281;
1282; HASWELL-NO-FMA-LABEL: v16f32_two_step:
1283; HASWELL-NO-FMA:       # %bb.0:
1284; HASWELL-NO-FMA-NEXT:    vrcpps %ymm0, %ymm2
1285; HASWELL-NO-FMA-NEXT:    vmulps %ymm2, %ymm0, %ymm3
1286; HASWELL-NO-FMA-NEXT:    vbroadcastss {{.*#+}} ymm4 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
1287; HASWELL-NO-FMA-NEXT:    vsubps %ymm3, %ymm4, %ymm3
1288; HASWELL-NO-FMA-NEXT:    vmulps %ymm3, %ymm2, %ymm3
1289; HASWELL-NO-FMA-NEXT:    vaddps %ymm3, %ymm2, %ymm2
1290; HASWELL-NO-FMA-NEXT:    vmulps %ymm2, %ymm0, %ymm0
1291; HASWELL-NO-FMA-NEXT:    vsubps %ymm0, %ymm4, %ymm0
1292; HASWELL-NO-FMA-NEXT:    vmulps %ymm0, %ymm2, %ymm0
1293; HASWELL-NO-FMA-NEXT:    vaddps %ymm0, %ymm2, %ymm0
1294; HASWELL-NO-FMA-NEXT:    vrcpps %ymm1, %ymm2
1295; HASWELL-NO-FMA-NEXT:    vmulps %ymm2, %ymm1, %ymm3
1296; HASWELL-NO-FMA-NEXT:    vsubps %ymm3, %ymm4, %ymm3
1297; HASWELL-NO-FMA-NEXT:    vmulps %ymm3, %ymm2, %ymm3
1298; HASWELL-NO-FMA-NEXT:    vaddps %ymm3, %ymm2, %ymm2
1299; HASWELL-NO-FMA-NEXT:    vmulps %ymm2, %ymm1, %ymm1
1300; HASWELL-NO-FMA-NEXT:    vsubps %ymm1, %ymm4, %ymm1
1301; HASWELL-NO-FMA-NEXT:    vmulps %ymm1, %ymm2, %ymm1
1302; HASWELL-NO-FMA-NEXT:    vaddps %ymm1, %ymm2, %ymm1
1303; HASWELL-NO-FMA-NEXT:    retq
1304;
1305; AVX512-LABEL: v16f32_two_step:
1306; AVX512:       # %bb.0:
1307; AVX512-NEXT:    vrcp14ps %zmm0, %zmm1
1308; AVX512-NEXT:    vbroadcastss {{.*#+}} zmm2 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0,1.0E+0]
1309; AVX512-NEXT:    vmovaps %zmm1, %zmm3
1310; AVX512-NEXT:    vfmsub213ps {{.*#+}} zmm3 = (zmm0 * zmm3) - zmm2
1311; AVX512-NEXT:    vfnmadd132ps {{.*#+}} zmm3 = -(zmm3 * zmm1) + zmm1
1312; AVX512-NEXT:    vfmsub213ps {{.*#+}} zmm0 = (zmm3 * zmm0) - zmm2
1313; AVX512-NEXT:    vfnmadd132ps {{.*#+}} zmm0 = -(zmm0 * zmm3) + zmm3
1314; AVX512-NEXT:    retq
1315  %div = fdiv fast <16 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, %x
1316  ret <16 x float> %div
1317}
1318
1319attributes #0 = { "unsafe-fp-math"="true" "reciprocal-estimates"="!divf,!vec-divf" }
1320attributes #1 = { "unsafe-fp-math"="true" "reciprocal-estimates"="divf,vec-divf" }
1321attributes #2 = { "unsafe-fp-math"="true" "reciprocal-estimates"="divf:2,vec-divf:2" }
1322
1323