mathvec/svml_s_sincosf16_core.s

   1 .macro WRAPPER_IMPL_SSE2 callee
   2         subq $40, %rsp
   3         movaps %xmm0, (%rsp)
   4         call \callee
   5         movss %xmm0, 16(%rsp)
   6         movss 4(%rsp), %xmm0
   7         call \callee
   8         movss %xmm0, 20(%rsp)
   9         movss 8(%rsp), %xmm0
  10         call \callee
  11         movss %xmm0, 24(%rsp)
  12         movss 12(%rsp), %xmm0
  13         call \callee
  14         movss 16(%rsp), %xmm3
  15         movss 20(%rsp), %xmm2
  16         movss 24(%rsp), %xmm1
  17         movss %xmm0, 28(%rsp)
  18         unpcklps %xmm1, %xmm3
  19         unpcklps %xmm0, %xmm2
  20         unpcklps %xmm2, %xmm3
  21         movaps %xmm3, %xmm0
  22         addq $40, %rsp
  23         ret
  24 .endm
  25 .macro WRAPPER_IMPL_SSE2_ff callee
  26         subq $56, %rsp
  27         movaps %xmm0, (%rsp)
  28         movaps %xmm1, 16(%rsp)
  29         call \callee
  30         movss %xmm0, 32(%rsp)
  31         movss 4(%rsp), %xmm0
  32         movss 20(%rsp), %xmm1
  33         call \callee
  34         movss %xmm0, 36(%rsp)
  35         movss 8(%rsp), %xmm0
  36         movss 24(%rsp), %xmm1
  37         call \callee
  38         movss %xmm0, 40(%rsp)
  39         movss 12(%rsp), %xmm0
  40         movss 28(%rsp), %xmm1
  41         call \callee
  42         movss 32(%rsp), %xmm3
  43         movss 36(%rsp), %xmm2
  44         movss 40(%rsp), %xmm1
  45         movss %xmm0, 44(%rsp)
  46         unpcklps %xmm1, %xmm3
  47         unpcklps %xmm0, %xmm2
  48         unpcklps %xmm2, %xmm3
  49         movaps %xmm3, %xmm0
  50         addq $56, %rsp
  51         ret
  52 .endm
  53 .macro WRAPPER_IMPL_SSE2_fFF callee
  54         pushq %rbp
  55         pushq %rbx
  56         movq %rdi, %rbp
  57         movq %rsi, %rbx
  58         subq $40, %rsp
  59         leaq 24(%rsp), %rsi
  60         leaq 28(%rsp), %rdi
  61         movaps %xmm0, (%rsp)
  62         call \callee
  63         leaq 24(%rsp), %rsi
  64         leaq 28(%rsp), %rdi
  65         movss 28(%rsp), %xmm0
  66         movss %xmm0, 0(%rbp)
  67         movaps (%rsp), %xmm1
  68         movss 24(%rsp), %xmm0
  69         movss %xmm0, (%rbx)
  70         movaps %xmm1, %xmm0
  71         shufps $85, %xmm1, %xmm0
  72         call \callee
  73         movss 28(%rsp), %xmm0
  74         leaq 24(%rsp), %rsi
  75         movss %xmm0, 4(%rbp)
  76         leaq 28(%rsp), %rdi
  77         movaps (%rsp), %xmm1
  78         movss 24(%rsp), %xmm0
  79         movss %xmm0, 4(%rbx)
  80         movaps %xmm1, %xmm0
  81         unpckhps %xmm1, %xmm0
  82         call \callee
  83         movaps (%rsp), %xmm1
  84         leaq 24(%rsp), %rsi
  85         leaq 28(%rsp), %rdi
  86         movss 28(%rsp), %xmm0
  87         shufps $255, %xmm1, %xmm1
  88         movss %xmm0, 8(%rbp)
  89         movss 24(%rsp), %xmm0
  90         movss %xmm0, 8(%rbx)
  91         movaps %xmm1, %xmm0
  92         call \callee
  93         movss 28(%rsp), %xmm0
  94         movss %xmm0, 12(%rbp)
  95         movss 24(%rsp), %xmm0
  96         movss %xmm0, 12(%rbx)
  97         addq $40, %rsp
  98         popq %rbx
  99         popq %rbp
 100         ret
 101 .endm
 102 .macro WRAPPER_IMPL_AVX callee
 103         pushq %rbp
 104         movq %rsp, %rbp
 105         andq $-32, %rsp
 106         subq $32, %rsp
 107         vextractf128 $1, %ymm0, (%rsp)
 108         vzeroupper
 109         call \callee
 110         vmovaps %xmm0, 16(%rsp)
 111         vmovaps (%rsp), %xmm0
 112         call \callee
 113         vmovaps %xmm0, %xmm1
 114         vmovaps 16(%rsp), %xmm0
 115         vinsertf128 $1, %xmm1, %ymm0, %ymm0
 116         movq %rbp, %rsp
 117         popq %rbp
 118         ret
 119 .endm
 120 .macro WRAPPER_IMPL_AVX_ff callee
 121         pushq %rbp
 122         movq %rsp, %rbp
 123         andq $-32, %rsp
 124         subq $64, %rsp
 125         vextractf128 $1, %ymm0, 16(%rsp)
 126         vextractf128 $1, %ymm1, (%rsp)
 127         vzeroupper
 128         call \callee
 129         vmovaps %xmm0, 32(%rsp)
 130         vmovaps 16(%rsp), %xmm0
 131         vmovaps (%rsp), %xmm1
 132         call \callee
 133         vmovaps %xmm0, %xmm1
 134         vmovaps 32(%rsp), %xmm0
 135         vinsertf128 $1, %xmm1, %ymm0, %ymm0
 136         movq %rbp, %rsp
 137         popq %rbp
 138         ret
 139 .endm
 140 .macro WRAPPER_IMPL_AVX_fFF callee
 141         pushq %rbp
 142         movq %rsp, %rbp
 143         andq $-32, %rsp
 144         pushq %r13
 145         pushq %r14
 146         subq $48, %rsp
 147         movq %rsi, %r14
 148         vmovaps %ymm0, (%rsp)
 149         movq %rdi, %r13
 150         vmovaps 16(%rsp), %xmm1
 151         vmovaps %xmm1, 32(%rsp)
 152         vzeroupper
 153         vmovaps (%rsp), %xmm0
 154         call \callee
 155         vmovaps 32(%rsp), %xmm0
 156         lea (%rsp), %rdi
 157         lea 16(%rsp), %rsi
 158         call \callee
 159         vmovaps (%rsp), %xmm0
 160         vmovaps 16(%rsp), %xmm1
 161         vmovaps %xmm0, 16(%r13)
 162         vmovaps %xmm1, 16(%r14)
 163         addq $48, %rsp
 164         popq %r14
 165         popq %r13
 166         movq %rbp, %rsp
 167         popq %rbp
 168         ret
 169 .endm
 170 .macro WRAPPER_IMPL_AVX512 callee
 171         pushq %rbp
 172         movq %rsp, %rbp
 173         andq $-64, %rsp
 174         subq $128, %rsp
 175         vmovups %zmm0, (%rsp)
 176         vmovupd (%rsp), %ymm0
 177         call \callee
 178         vmovupd %ymm0, 64(%rsp)
 179         vmovupd 32(%rsp), %ymm0
 180         call \callee
 181         vmovupd %ymm0, 96(%rsp)
 182         vmovups 64(%rsp), %zmm0
 183         movq %rbp, %rsp
 184         popq %rbp
 185         ret
 186 .endm
 187 .macro WRAPPER_IMPL_AVX512_ff callee
 188         pushq %rbp
 189         movq %rsp, %rbp
 190         andq $-64, %rsp
 191         subq $192, %rsp
 192         vmovups %zmm0, (%rsp)
 193         vmovups %zmm1, 64(%rsp)
 194         vmovups (%rsp), %ymm0
 195         vmovups 64(%rsp), %ymm1
 196         call \callee
 197         vmovups %ymm0, 128(%rsp)
 198         vmovups 32(%rsp), %ymm0
 199         vmovups 96(%rsp), %ymm1
 200         call \callee
 201         vmovups %ymm0, 160(%rsp)
 202         vmovups 128(%rsp), %zmm0
 203         movq %rbp, %rsp
 204         popq %rbp
 205         ret
 206 .endm
 207 .macro WRAPPER_IMPL_AVX512_fFF callee
 208         pushq %rbp
 209         movq %rsp, %rbp
 210         andq $-64, %rsp
 211         pushq %r12
 212         pushq %r13
 213         subq $176, %rsp
 214         movq %rsi, %r13
 215         vmovaps %zmm0, (%rsp)
 216         movq %rdi, %r12
 217         vmovaps (%rsp), %ymm0
 218         call \callee
 219         vmovaps 32(%rsp), %ymm0
 220         lea 64(%rsp), %rdi
 221         lea 96(%rsp), %rsi
 222         call \callee
 223         vmovaps 64(%rsp), %ymm0
 224         vmovaps 96(%rsp), %ymm1
 225         vmovaps %ymm0, 32(%r12)
 226         vmovaps %ymm1, 32(%r13)
 227         addq $176, %rsp
 228         popq %r13
 229         popq %r12
 230         movq %rbp, %rsp
 231         popq %rbp
 232         ret
 233 .endm
 234  .text
 235 .globl _ZGVeN16vl4l4_sincosf
 236  .type _ZGVeN16vl4l4_sincosf,@function
 237  .align 1<<4
 238  _ZGVeN16vl4l4_sincosf:
 239
 240
 241 WRAPPER_IMPL_AVX512_fFF _ZGVdN8vl4l4_sincosf
 242  .size _ZGVeN16vl4l4_sincosf,.-_ZGVeN16vl4l4_sincosf
 243
 244 .macro WRAPPER_IMPL_AVX512_fFF_vvv callee
 245         pushq %rbp
 246         movq %rsp, %rbp
 247         andq $-64, %rsp
 248         subq $448, %rsp
 249         vmovups %zmm0, 384(%rsp)
 250         lea (%rsp), %rdi
 251         vmovups %zmm1, 128(%rdi)
 252         vmovups %zmm2, 192(%rdi)
 253         vmovups %zmm3, 256(%rdi)
 254         vmovups %zmm4, 320(%rdi)
 255         lea 64(%rsp), %rsi
 256         call \callee
 257         vmovdqu 416(%rsp), %ymm0
 258         lea 32(%rsp), %rdi
 259         lea 96(%rsp), %rsi
 260         call \callee
 261         movq 128(%rsp), %rdx
 262         movq 136(%rsp), %rsi
 263         movq 144(%rsp), %r8
 264         movq 152(%rsp), %r10
 265         movl (%rsp), %eax
 266         movl 4(%rsp), %ecx
 267         movl 8(%rsp), %edi
 268         movl 12(%rsp), %r9d
 269         movl %eax, (%rdx)
 270         movl %ecx, (%rsi)
 271         movq 160(%rsp), %rax
 272         movq 168(%rsp), %rcx
 273         movl %edi, (%r8)
 274         movl %r9d, (%r10)
 275         movq 176(%rsp), %rdi
 276         movq 184(%rsp), %r9
 277         movl 16(%rsp), %r11d
 278         movl 20(%rsp), %edx
 279         movl 24(%rsp), %esi
 280         movl 28(%rsp), %r8d
 281         movl %r11d, (%rax)
 282         movl %edx, (%rcx)
 283         movq 192(%rsp), %r11
 284         movq 200(%rsp), %rdx
 285         movl %esi, (%rdi)
 286         movl %r8d, (%r9)
 287         movq 208(%rsp), %rsi
 288         movq 216(%rsp), %r8
 289         movl 32(%rsp), %r10d
 290         movl 36(%rsp), %eax
 291         movl 40(%rsp), %ecx
 292         movl 44(%rsp), %edi
 293         movl %r10d, (%r11)
 294         movl %eax, (%rdx)
 295         movq 224(%rsp), %r10
 296         movq 232(%rsp), %rax
 297         movl %ecx, (%rsi)
 298         movl %edi, (%r8)
 299         movq 240(%rsp), %rcx
 300         movq 248(%rsp), %rdi
 301         movl 48(%rsp), %r9d
 302         movl 52(%rsp), %r11d
 303         movl 56(%rsp), %edx
 304         movl 60(%rsp), %esi
 305         movl %r9d, (%r10)
 306         movl %r11d, (%rax)
 307         movq 256(%rsp), %r9
 308         movq 264(%rsp), %r11
 309         movl %edx, (%rcx)
 310         movl %esi, (%rdi)
 311         movq 272(%rsp), %rdx
 312         movq 280(%rsp), %rsi
 313         movl 64(%rsp), %r8d
 314         movl 68(%rsp), %r10d
 315         movl 72(%rsp), %eax
 316         movl 76(%rsp), %ecx
 317         movl %r8d, (%r9)
 318         movl %r10d, (%r11)
 319         movq 288(%rsp), %r8
 320         movq 296(%rsp), %r10
 321         movl %eax, (%rdx)
 322         movl %ecx, (%rsi)
 323         movq 304(%rsp), %rax
 324         movq 312(%rsp), %rcx
 325         movl 80(%rsp), %edi
 326         movl 84(%rsp), %r9d
 327         movl 88(%rsp), %r11d
 328         movl 92(%rsp), %edx
 329         movl %edi, (%r8)
 330         movl %r9d, (%r10)
 331         movq 320(%rsp), %rdi
 332         movq 328(%rsp), %r9
 333         movl %r11d, (%rax)
 334         movl %edx, (%rcx)
 335         movq 336(%rsp), %r11
 336         movq 344(%rsp), %rdx
 337         movl 96(%rsp), %esi
 338         movl 100(%rsp), %r8d
 339         movl 104(%rsp), %r10d
 340         movl 108(%rsp), %eax
 341         movl %esi, (%rdi)
 342         movl %r8d, (%r9)
 343         movq 352(%rsp), %rsi
 344         movq 360(%rsp), %r8
 345         movl %r10d, (%r11)
 346         movl %eax, (%rdx)
 347         movq 368(%rsp), %r10
 348         movq 376(%rsp), %rax
 349         movl 112(%rsp), %ecx
 350         movl 116(%rsp), %edi
 351         movl 120(%rsp), %r9d
 352         movl 124(%rsp), %r11d
 353         movl %ecx, (%rsi)
 354         movl %edi, (%r8)
 355         movl %r9d, (%r10)
 356         movl %r11d, (%rax)
 357         movq %rbp, %rsp
 358         popq %rbp
 359         ret
 360 .endm
 361 .globl _ZGVeN16vvv_sincosf
 362  .type _ZGVeN16vvv_sincosf,@function
 363  .align 1<<4
 364  _ZGVeN16vvv_sincosf:
 365
 366
 367 WRAPPER_IMPL_AVX512_fFF_vvv _ZGVdN8vl4l4_sincosf
 368  .size _ZGVeN16vvv_sincosf,.-_ZGVeN16vvv_sincosf
 369