Lines Matching full:fast

4 …UN: llc < %s -mtriple=x86_64-unknown -mattr=+ssse3,fast-hops | FileCheck %s --check-prefix=SSSE3-F…
6 … RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx,fast-hops   | FileCheck %s --check-prefix=AVX1-…
32 ; SSSE3-FAST-LABEL: PR37890_v4f32:
33 ; SSSE3-FAST:       # %bb.0:
34 ; SSSE3-FAST-NEXT:    haddps %xmm0, %xmm0
35 ; SSSE3-FAST-NEXT:    haddps %xmm0, %xmm0
36 ; SSSE3-FAST-NEXT:    retq
46 ; AVX1-FAST-LABEL: PR37890_v4f32:
47 ; AVX1-FAST:       # %bb.0:
48 ; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
49 ; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
50 ; AVX1-FAST-NEXT:    retq
61   %sum0 = fadd fast <2 x float> %lo0, %hi0
63   %sum1 = fadd fast <2 x float> %sum0, %hi1
87 ; SSSE3-FAST-LABEL: PR37890_v4f64:
88 ; SSSE3-FAST:       # %bb.0:
89 ; SSSE3-FAST-NEXT:    addpd %xmm1, %xmm0
90 ; SSSE3-FAST-NEXT:    haddpd %xmm0, %xmm0
91 ; SSSE3-FAST-NEXT:    retq
102 ; AVX1-FAST-LABEL: PR37890_v4f64:
103 ; AVX1-FAST:       # %bb.0:
104 ; AVX1-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm1
105 ; AVX1-FAST-NEXT:    vhaddpd %xmm0, %xmm1, %xmm0
106 ; AVX1-FAST-NEXT:    vhaddpd %xmm0, %xmm0, %xmm0
107 ; AVX1-FAST-NEXT:    vzeroupper
108 ; AVX1-FAST-NEXT:    retq
120   %sum0 = fadd fast <2 x double> %lo0, %hi0
122   %sum1 = fadd fast <2 x double> %sum0, %hi1
150 ; SSSE3-FAST-LABEL: PR37890_v8f32:
151 ; SSSE3-FAST:       # %bb.0:
152 ; SSSE3-FAST-NEXT:    addps %xmm1, %xmm0
153 ; SSSE3-FAST-NEXT:    haddps %xmm0, %xmm0
154 ; SSSE3-FAST-NEXT:    haddps %xmm0, %xmm0
155 ; SSSE3-FAST-NEXT:    retq
168 ; AVX1-FAST-LABEL: PR37890_v8f32:
169 ; AVX1-FAST:       # %bb.0:
170 ; AVX1-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm1
171 ; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm1, %xmm0
172 ; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
173 ; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
174 ; AVX1-FAST-NEXT:    vzeroupper
175 ; AVX1-FAST-NEXT:    retq
189   %sum0 = fadd fast <4 x float> %lo0, %hi0
192   %sum1 = fadd fast <2 x float> %lo1, %hi1
194   %sum2 = fadd fast <2 x float> %sum1, %hi2
220 ; SSSE3-FAST-LABEL: PR37890_v8f64:
221 ; SSSE3-FAST:       # %bb.0:
222 ; SSSE3-FAST-NEXT:    addpd %xmm3, %xmm1
223 ; SSSE3-FAST-NEXT:    addpd %xmm2, %xmm1
224 ; SSSE3-FAST-NEXT:    addpd %xmm0, %xmm1
225 ; SSSE3-FAST-NEXT:    haddpd %xmm1, %xmm1
226 ; SSSE3-FAST-NEXT:    movapd %xmm1, %xmm0
227 ; SSSE3-FAST-NEXT:    retq
239 ; AVX1-FAST-LABEL: PR37890_v8f64:
240 ; AVX1-FAST:       # %bb.0:
241 ; AVX1-FAST-NEXT:    vaddpd %ymm1, %ymm0, %ymm0
242 ; AVX1-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm1
243 ; AVX1-FAST-NEXT:    vhaddpd %xmm0, %xmm1, %xmm0
244 ; AVX1-FAST-NEXT:    vhaddpd %xmm0, %xmm0, %xmm0
245 ; AVX1-FAST-NEXT:    vzeroupper
246 ; AVX1-FAST-NEXT:    retq
259   %sum0 = fadd fast <4 x double> %lo0, %hi0
262   %sum1 = fadd fast <2 x double> %lo1, %hi1
264   %sum2 = fadd fast <2 x double> %sum1, %hi2
295 ; SSSE3-FAST-LABEL: PR37890_v16f32:
296 ; SSSE3-FAST:       # %bb.0:
297 ; SSSE3-FAST-NEXT:    addps %xmm3, %xmm1
298 ; SSSE3-FAST-NEXT:    addps %xmm2, %xmm1
299 ; SSSE3-FAST-NEXT:    addps %xmm0, %xmm1
300 ; SSSE3-FAST-NEXT:    movaps %xmm1, %xmm0
301 ; SSSE3-FAST-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
302 ; SSSE3-FAST-NEXT:    addps %xmm1, %xmm0
303 ; SSSE3-FAST-NEXT:    haddps %xmm0, %xmm0
304 ; SSSE3-FAST-NEXT:    retq
318 ; AVX1-FAST-LABEL: PR37890_v16f32:
319 ; AVX1-FAST:       # %bb.0:
320 ; AVX1-FAST-NEXT:    vaddps %ymm1, %ymm0, %ymm0
321 ; AVX1-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm1
322 ; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm1, %xmm0
323 ; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
324 ; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
325 ; AVX1-FAST-NEXT:    vzeroupper
326 ; AVX1-FAST-NEXT:    retq
341   %sum0 = fadd fast <8 x float> %lo0, %hi0
344   %sum1 = fadd fast <4 x float> %lo1, %hi1
347   %sum2 = fadd fast <2 x float> %lo2, %hi2
349   %sum3 = fadd fast <2 x float> %sum2, %hi3