865 lines · plain
1 .text2 .file "matmul.c"3 .globl init_array # -- Begin function init_array4 .p2align 4, 0x905 .type init_array,@function6init_array: # @init_array7 .cfi_startproc8# %bb.0: # %entry9 pushq %rbp10 .cfi_def_cfa_offset 1611 .cfi_offset %rbp, -1612 movq %rsp, %rbp13 .cfi_def_cfa_register %rbp14 pushq %rbx15 pushq %rax16 .cfi_offset %rbx, -2417 leaq init_array_polly_subfn(%rip), %rdi18 leaq -16(%rbp), %rbx19 xorl %edx, %edx20 xorl %ecx, %ecx21 movl $1536, %r8d # imm = 0x60022 movl $1, %r9d23 movq %rbx, %rsi24 callq GOMP_parallel_loop_runtime_start@PLT25 movq %rbx, %rdi26 callq init_array_polly_subfn27 callq GOMP_parallel_end@PLT28 addq $8, %rsp29 popq %rbx30 popq %rbp31 .cfi_def_cfa %rsp, 832 retq33.Lfunc_end0:34 .size init_array, .Lfunc_end0-init_array35 .cfi_endproc36 # -- End function37 .globl print_array # -- Begin function print_array38 .p2align 4, 0x9039 .type print_array,@function40print_array: # @print_array41 .cfi_startproc42# %bb.0: # %entry43 pushq %rbp44 .cfi_def_cfa_offset 1645 .cfi_offset %rbp, -1646 movq %rsp, %rbp47 .cfi_def_cfa_register %rbp48 pushq %r1549 pushq %r1450 pushq %r1351 pushq %r1252 pushq %rbx53 pushq %rax54 .cfi_offset %rbx, -5655 .cfi_offset %r12, -4856 .cfi_offset %r13, -4057 .cfi_offset %r14, -3258 .cfi_offset %r15, -2459 leaq C(%rip), %r1360 xorl %eax, %eax61 movl $3435973837, %r12d # imm = 0xCCCCCCCD62 leaq .L.str(%rip), %r1463 .p2align 4, 0x9064.LBB1_1: # %for.cond1.preheader65 # =>This Loop Header: Depth=166 # Child Loop BB1_2 Depth 267 movq %rax, -48(%rbp) # 8-byte Spill68 movq stdout(%rip), %rsi69 xorl %ebx, %ebx70 .p2align 4, 0x9071.LBB1_2: # %for.body372 # Parent Loop BB1_1 Depth=173 # => This Inner Loop Header: Depth=274 movl %ebx, %eax75 imulq %r12, %rax76 shrq $38, %rax77 leal (%rax,%rax,4), %r15d78 shll $4, %r15d79 addl $79, %r15d80 movss (%r13,%rbx,4), %xmm0 # xmm0 = mem[0],zero,zero,zero81 cvtss2sd %xmm0, %xmm082 movb $1, %al83 movq %rsi, %rdi84 movq %r14, %rsi85 callq fprintf86 cmpl %ebx, %r15d87 jne .LBB1_488# %bb.3: # %if.then89 # in Loop: Header=BB1_2 Depth=290 movq stdout(%rip), %rsi91 movl $10, %edi92 callq fputc@PLT93.LBB1_4: # %for.inc94 # in Loop: Header=BB1_2 Depth=295 addq $1, %rbx96 movq stdout(%rip), %rsi97 cmpq $1536, %rbx # imm = 0x60098 jne .LBB1_299# %bb.5: # %for.end100 # in Loop: Header=BB1_1 Depth=1101 movl $10, %edi102 callq fputc@PLT103 movq -48(%rbp), %rax # 8-byte Reload104 addq $1, %rax105 addq $6144, %r13 # imm = 0x1800106 cmpq $1536, %rax # imm = 0x600107 jne .LBB1_1108# %bb.6: # %for.end12109 addq $8, %rsp110 popq %rbx111 popq %r12112 popq %r13113 popq %r14114 popq %r15115 popq %rbp116 .cfi_def_cfa %rsp, 8117 retq118.Lfunc_end1:119 .size print_array, .Lfunc_end1-print_array120 .cfi_endproc121 # -- End function122 .globl main # -- Begin function main123 .p2align 4, 0x90124 .type main,@function125main: # @main126 .cfi_startproc127# %bb.0: # %entry128 pushq %rbp129 .cfi_def_cfa_offset 16130 .cfi_offset %rbp, -16131 movq %rsp, %rbp132 .cfi_def_cfa_register %rbp133 pushq %rbx134 pushq %rax135 .cfi_offset %rbx, -24136 callq init_array137 leaq main_polly_subfn(%rip), %rdi138 leaq -16(%rbp), %rbx139 xorl %edx, %edx140 xorl %ecx, %ecx141 movl $1536, %r8d # imm = 0x600142 movl $1, %r9d143 movq %rbx, %rsi144 callq GOMP_parallel_loop_runtime_start@PLT145 movq %rbx, %rdi146 callq main_polly_subfn147 callq GOMP_parallel_end@PLT148 leaq main_polly_subfn_1(%rip), %rdi149 xorl %edx, %edx150 xorl %ecx, %ecx151 movl $1536, %r8d # imm = 0x600152 movl $64, %r9d153 movq %rbx, %rsi154 callq GOMP_parallel_loop_runtime_start@PLT155 movq %rbx, %rdi156 callq main_polly_subfn_1157 callq GOMP_parallel_end@PLT158 xorl %eax, %eax159 addq $8, %rsp160 popq %rbx161 popq %rbp162 .cfi_def_cfa %rsp, 8163 retq164.Lfunc_end2:165 .size main, .Lfunc_end2-main166 .cfi_endproc167 # -- End function168 .section .rodata.cst8,"aM",@progbits,8169 .p2align 3 # -- Begin function init_array_polly_subfn170.LCPI3_0:171 .quad 4602678819172646912 # double 0.5172 .text173 .p2align 4, 0x90174 .type init_array_polly_subfn,@function175init_array_polly_subfn: # @init_array_polly_subfn176 .cfi_startproc177# %bb.0: # %polly.par.setup178 pushq %r15179 .cfi_def_cfa_offset 16180 pushq %r14181 .cfi_def_cfa_offset 24182 pushq %r13183 .cfi_def_cfa_offset 32184 pushq %r12185 .cfi_def_cfa_offset 40186 pushq %rbx187 .cfi_def_cfa_offset 48188 subq $16, %rsp189 .cfi_def_cfa_offset 64190 .cfi_offset %rbx, -48191 .cfi_offset %r12, -40192 .cfi_offset %r13, -32193 .cfi_offset %r14, -24194 .cfi_offset %r15, -16195 leaq 8(%rsp), %rdi196 movq %rsp, %rsi197 callq GOMP_loop_runtime_next@PLT198 testb %al, %al199 je .LBB3_2200# %bb.1:201 leaq B(%rip), %r15202 leaq A(%rip), %r12203 movsd .LCPI3_0(%rip), %xmm1 # xmm1 = mem[0],zero204 leaq 8(%rsp), %r14205 movq %rsp, %r13206 .p2align 4, 0x90207.LBB3_4: # %polly.par.loadIVBounds208 # =>This Loop Header: Depth=1209 # Child Loop BB3_5 Depth 2210 # Child Loop BB3_6 Depth 3211 movq 8(%rsp), %rax212 movq (%rsp), %r8213 decq %r8214 movq %rax, %rdx215 shlq $11, %rdx216 leaq (%rdx,%rdx,2), %rdx217 leaq (%r15,%rdx), %rsi218 addq %r12, %rdx219 .p2align 4, 0x90220.LBB3_5: # %polly.loop_header221 # Parent Loop BB3_4 Depth=1222 # => This Loop Header: Depth=2223 # Child Loop BB3_6 Depth 3224 movq $-6144, %rdi # imm = 0xE800225 xorl %ecx, %ecx226 .p2align 4, 0x90227.LBB3_6: # %polly.loop_header2228 # Parent Loop BB3_4 Depth=1229 # Parent Loop BB3_5 Depth=2230 # => This Inner Loop Header: Depth=3231 movl %ecx, %ebx232 andl $1023, %ebx # imm = 0x3FF233 incl %ebx234 xorps %xmm0, %xmm0235 cvtsi2sdl %ebx, %xmm0236 mulsd %xmm1, %xmm0237 cvtsd2ss %xmm0, %xmm0238 movss %xmm0, 6144(%rdx,%rdi)239 movss %xmm0, 6144(%rsi,%rdi)240 addl %eax, %ecx241 addq $4, %rdi242 jne .LBB3_6243# %bb.7: # %polly.loop_exit4244 # in Loop: Header=BB3_5 Depth=2245 addq $6144, %rsi # imm = 0x1800246 addq $6144, %rdx # imm = 0x1800247 cmpq %r8, %rax248 leaq 1(%rax), %rax249 jl .LBB3_5250# %bb.3: # %polly.par.checkNext.loopexit251 # in Loop: Header=BB3_4 Depth=1252 movq %r14, %rdi253 movq %r13, %rsi254 callq GOMP_loop_runtime_next@PLT255 movsd .LCPI3_0(%rip), %xmm1 # xmm1 = mem[0],zero256 testb %al, %al257 jne .LBB3_4258.LBB3_2: # %polly.par.exit259 callq GOMP_loop_end_nowait@PLT260 addq $16, %rsp261 .cfi_def_cfa_offset 48262 popq %rbx263 .cfi_def_cfa_offset 40264 popq %r12265 .cfi_def_cfa_offset 32266 popq %r13267 .cfi_def_cfa_offset 24268 popq %r14269 .cfi_def_cfa_offset 16270 popq %r15271 .cfi_def_cfa_offset 8272 retq273.Lfunc_end3:274 .size init_array_polly_subfn, .Lfunc_end3-init_array_polly_subfn275 .cfi_endproc276 # -- End function277 .p2align 4, 0x90 # -- Begin function main_polly_subfn278 .type main_polly_subfn,@function279main_polly_subfn: # @main_polly_subfn280 .cfi_startproc281# %bb.0: # %polly.par.setup282 pushq %r15283 .cfi_def_cfa_offset 16284 pushq %r14285 .cfi_def_cfa_offset 24286 pushq %rbx287 .cfi_def_cfa_offset 32288 subq $16, %rsp289 .cfi_def_cfa_offset 48290 .cfi_offset %rbx, -32291 .cfi_offset %r14, -24292 .cfi_offset %r15, -16293 leaq 8(%rsp), %rdi294 movq %rsp, %rsi295 callq GOMP_loop_runtime_next@PLT296 testb %al, %al297 je .LBB4_3298# %bb.1:299 leaq C(%rip), %r15300 leaq 8(%rsp), %r14301 movq %rsp, %rbx302 .p2align 4, 0x90303.LBB4_2: # %polly.par.loadIVBounds304 # =>This Inner Loop Header: Depth=1305 movq 8(%rsp), %rax306 movq (%rsp), %rcx307 decq %rcx308 leaq (%rax,%rax,2), %rdi309 shlq $11, %rdi310 addq %r15, %rdi311 cmpq %rcx, %rax312 cmovgeq %rax, %rcx313 incq %rcx314 subq %rax, %rcx315 shlq $11, %rcx316 leaq (%rcx,%rcx,2), %rdx317 xorl %esi, %esi318 callq memset@PLT319 movq %r14, %rdi320 movq %rbx, %rsi321 callq GOMP_loop_runtime_next@PLT322 testb %al, %al323 jne .LBB4_2324.LBB4_3: # %polly.par.exit325 callq GOMP_loop_end_nowait@PLT326 addq $16, %rsp327 .cfi_def_cfa_offset 32328 popq %rbx329 .cfi_def_cfa_offset 24330 popq %r14331 .cfi_def_cfa_offset 16332 popq %r15333 .cfi_def_cfa_offset 8334 retq335.Lfunc_end4:336 .size main_polly_subfn, .Lfunc_end4-main_polly_subfn337 .cfi_endproc338 # -- End function339 .p2align 4, 0x90 # -- Begin function main_polly_subfn_1340 .type main_polly_subfn_1,@function341main_polly_subfn_1: # @main_polly_subfn_1342 .cfi_startproc343# %bb.0: # %polly.par.setup344 pushq %rbp345 .cfi_def_cfa_offset 16346 pushq %r15347 .cfi_def_cfa_offset 24348 pushq %r14349 .cfi_def_cfa_offset 32350 pushq %r13351 .cfi_def_cfa_offset 40352 pushq %r12353 .cfi_def_cfa_offset 48354 pushq %rbx355 .cfi_def_cfa_offset 56356 subq $296, %rsp # imm = 0x128357 .cfi_def_cfa_offset 352358 .cfi_offset %rbx, -56359 .cfi_offset %r12, -48360 .cfi_offset %r13, -40361 .cfi_offset %r14, -32362 .cfi_offset %r15, -24363 .cfi_offset %rbp, -16364 jmp .LBB5_1365 .p2align 4, 0x90366.LBB5_2: # %polly.par.loadIVBounds367 # in Loop: Header=BB5_1 Depth=1368 movq 40(%rsp), %rdx369 movq 32(%rsp), %rax370 decq %rax371 movq %rax, 136(%rsp) # 8-byte Spill372 leaq (%rdx,%rdx,2), %rcx373 shlq $11, %rcx374 leaq A(%rip), %rax375 addq %rax, %rcx376 movq %rcx, 24(%rsp) # 8-byte Spill377 .p2align 4, 0x90378.LBB5_3: # %polly.loop_header379 # Parent Loop BB5_1 Depth=1380 # => This Loop Header: Depth=2381 # Child Loop BB5_4 Depth 3382 # Child Loop BB5_5 Depth 4383 # Child Loop BB5_6 Depth 5384 # Child Loop BB5_7 Depth 6385 leaq 63(%rdx), %rsi386 leaq B+192(%rip), %r14387 xorl %ecx, %ecx388 xorl %eax, %eax389 movq %rdx, 168(%rsp) # 8-byte Spill390 .p2align 4, 0x90391.LBB5_4: # %polly.loop_header2392 # Parent Loop BB5_1 Depth=1393 # Parent Loop BB5_3 Depth=2394 # => This Loop Header: Depth=3395 # Child Loop BB5_5 Depth 4396 # Child Loop BB5_6 Depth 5397 # Child Loop BB5_7 Depth 6398 movq %rax, 144(%rsp) # 8-byte Spill399 movq %rcx, 152(%rsp) # 8-byte Spill400 shlq $6, %rcx401 leaq 16(%rcx), %rdi402 leaq 32(%rcx), %rbp403 leaq 48(%rcx), %r15404 movq 24(%rsp), %r9 # 8-byte Reload405 movq %r14, 160(%rsp) # 8-byte Spill406 xorl %eax, %eax407 .p2align 4, 0x90408.LBB5_5: # %polly.loop_header8409 # Parent Loop BB5_1 Depth=1410 # Parent Loop BB5_3 Depth=2411 # Parent Loop BB5_4 Depth=3412 # => This Loop Header: Depth=4413 # Child Loop BB5_6 Depth 5414 # Child Loop BB5_7 Depth 6415 movq %rax, 176(%rsp) # 8-byte Spill416 movq %r9, 184(%rsp) # 8-byte Spill417 movq %rdx, %rax418 .p2align 4, 0x90419.LBB5_6: # %polly.loop_header14420 # Parent Loop BB5_1 Depth=1421 # Parent Loop BB5_3 Depth=2422 # Parent Loop BB5_4 Depth=3423 # Parent Loop BB5_5 Depth=4424 # => This Loop Header: Depth=5425 # Child Loop BB5_7 Depth 6426 leaq (%rax,%rax,2), %rbx427 shlq $11, %rbx428 leaq C(%rip), %rdx429 addq %rdx, %rbx430 leaq (%rbx,%rcx,4), %r8431 leaq (%rbx,%rdi,4), %rdx432 leaq (%rbx,%rbp,4), %r13433 leaq (%rbx,%r15,4), %r10434 movups (%rbx,%rcx,4), %xmm8435 movups 16(%rbx,%rcx,4), %xmm0436 movaps %xmm0, 96(%rsp) # 16-byte Spill437 movups 32(%rbx,%rcx,4), %xmm6438 movups 48(%rbx,%rcx,4), %xmm1439 movups (%rbx,%rdi,4), %xmm15440 movups 16(%rbx,%rdi,4), %xmm0441 movaps %xmm0, (%rsp) # 16-byte Spill442 movups 32(%rbx,%rdi,4), %xmm0443 movaps %xmm0, 48(%rsp) # 16-byte Spill444 movups 48(%rbx,%rdi,4), %xmm0445 movaps %xmm0, 64(%rsp) # 16-byte Spill446 movups (%rbx,%rbp,4), %xmm11447 movups 16(%rbx,%rbp,4), %xmm0448 movaps %xmm0, 112(%rsp) # 16-byte Spill449 movups 32(%rbx,%rbp,4), %xmm12450 movups 48(%rbx,%rbp,4), %xmm0451 movaps %xmm0, 80(%rsp) # 16-byte Spill452 movups (%rbx,%r15,4), %xmm9453 movups 16(%rbx,%r15,4), %xmm13454 movups 32(%rbx,%r15,4), %xmm2455 movups 48(%rbx,%r15,4), %xmm3456 movq $-256, %r12457 movq %r14, %r11458 .p2align 4, 0x90459.LBB5_7: # %vector.ph460 # Parent Loop BB5_1 Depth=1461 # Parent Loop BB5_3 Depth=2462 # Parent Loop BB5_4 Depth=3463 # Parent Loop BB5_5 Depth=4464 # Parent Loop BB5_6 Depth=5465 # => This Inner Loop Header: Depth=6466 movaps %xmm12, 208(%rsp) # 16-byte Spill467 movaps %xmm2, 224(%rsp) # 16-byte Spill468 movaps %xmm3, 240(%rsp) # 16-byte Spill469 movaps %xmm8, %xmm10470 movaps 96(%rsp), %xmm7 # 16-byte Reload471 unpcklps %xmm7, %xmm10 # xmm10 = xmm10[0],xmm7[0],xmm10[1],xmm7[1]472 movaps %xmm1, %xmm4473 shufps $0, %xmm6, %xmm4 # xmm4 = xmm4[0,0],xmm6[0,0]474 shufps $36, %xmm4, %xmm10 # xmm10 = xmm10[0,1],xmm4[2,0]475 movaps %xmm7, %xmm5476 shufps $17, %xmm8, %xmm5 # xmm5 = xmm5[1,0],xmm8[1,0]477 movaps %xmm6, %xmm4478 unpcklps %xmm1, %xmm4 # xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1]479 shufps $226, %xmm4, %xmm5 # xmm5 = xmm5[2,0],xmm4[2,3]480 movaps %xmm8, %xmm12481 unpckhps %xmm7, %xmm12 # xmm12 = xmm12[2],xmm7[2],xmm12[3],xmm7[3]482 movaps %xmm1, %xmm4483 shufps $34, %xmm6, %xmm4 # xmm4 = xmm4[2,0],xmm6[2,0]484 shufps $36, %xmm4, %xmm12 # xmm12 = xmm12[0,1],xmm4[2,0]485 shufps $51, %xmm8, %xmm7 # xmm7 = xmm7[3,0],xmm8[3,0]486 unpckhps %xmm1, %xmm6 # xmm6 = xmm6[2],xmm1[2],xmm6[3],xmm1[3]487 shufps $226, %xmm6, %xmm7 # xmm7 = xmm7[2,0],xmm6[2,3]488 movaps -160(%r11), %xmm0489 movaps -144(%r11), %xmm1490 movaps %xmm1, %xmm6491 shufps $0, %xmm0, %xmm6 # xmm6 = xmm6[0,0],xmm0[0,0]492 movaps -192(%r11), %xmm3493 movaps -176(%r11), %xmm4494 movaps %xmm3, %xmm8495 unpcklps %xmm4, %xmm8 # xmm8 = xmm8[0],xmm4[0],xmm8[1],xmm4[1]496 shufps $36, %xmm6, %xmm8 # xmm8 = xmm8[0,1],xmm6[2,0]497 movaps %xmm0, %xmm2498 unpcklps %xmm1, %xmm2 # xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]499 movaps %xmm4, %xmm6500 shufps $17, %xmm3, %xmm6 # xmm6 = xmm6[1,0],xmm3[1,0]501 shufps $226, %xmm2, %xmm6 # xmm6 = xmm6[2,0],xmm2[2,3]502 movaps %xmm1, %xmm2503 shufps $34, %xmm0, %xmm2 # xmm2 = xmm2[2,0],xmm0[2,0]504 movaps %xmm3, %xmm14505 unpckhps %xmm4, %xmm14 # xmm14 = xmm14[2],xmm4[2],xmm14[3],xmm4[3]506 shufps $36, %xmm2, %xmm14 # xmm14 = xmm14[0,1],xmm2[2,0]507 unpckhps %xmm1, %xmm0 # xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]508 shufps $51, %xmm3, %xmm4 # xmm4 = xmm4[3,0],xmm3[3,0]509 shufps $226, %xmm0, %xmm4 # xmm4 = xmm4[2,0],xmm0[2,3]510 movss 256(%r9,%r12), %xmm0 # xmm0 = mem[0],zero,zero,zero511 shufps $0, %xmm0, %xmm0 # xmm0 = xmm0[0,0,0,0]512 mulps %xmm0, %xmm8513 addps %xmm10, %xmm8514 mulps %xmm0, %xmm6515 addps %xmm5, %xmm6516 mulps %xmm0, %xmm14517 addps %xmm12, %xmm14518 mulps %xmm0, %xmm4519 movaps %xmm0, %xmm5520 addps %xmm7, %xmm4521 movaps %xmm14, %xmm0522 unpckhps %xmm4, %xmm0 # xmm0 = xmm0[2],xmm4[2],xmm0[3],xmm4[3]523 movaps %xmm6, %xmm1524 shufps $51, %xmm8, %xmm1 # xmm1 = xmm1[3,0],xmm8[3,0]525 shufps $226, %xmm0, %xmm1 # xmm1 = xmm1[2,0],xmm0[2,3]526 movaps %xmm1, 272(%rsp) # 16-byte Spill527 movaps %xmm4, %xmm0528 shufps $34, %xmm14, %xmm0 # xmm0 = xmm0[2,0],xmm14[2,0]529 movaps %xmm8, %xmm1530 unpckhps %xmm6, %xmm1 # xmm1 = xmm1[2],xmm6[2],xmm1[3],xmm6[3]531 shufps $36, %xmm0, %xmm1 # xmm1 = xmm1[0,1],xmm0[2,0]532 movaps %xmm1, 256(%rsp) # 16-byte Spill533 movaps %xmm14, %xmm0534 unpcklps %xmm4, %xmm0 # xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]535 movaps %xmm6, %xmm1536 shufps $17, %xmm8, %xmm1 # xmm1 = xmm1[1,0],xmm8[1,0]537 shufps $226, %xmm0, %xmm1 # xmm1 = xmm1[2,0],xmm0[2,3]538 movaps %xmm1, 96(%rsp) # 16-byte Spill539 shufps $0, %xmm14, %xmm4 # xmm4 = xmm4[0,0],xmm14[0,0]540 unpcklps %xmm6, %xmm8 # xmm8 = xmm8[0],xmm6[0],xmm8[1],xmm6[1]541 shufps $36, %xmm4, %xmm8 # xmm8 = xmm8[0,1],xmm4[2,0]542 movaps %xmm15, %xmm14543 movaps (%rsp), %xmm4 # 16-byte Reload544 unpcklps %xmm4, %xmm14 # xmm14 = xmm14[0],xmm4[0],xmm14[1],xmm4[1]545 movaps 64(%rsp), %xmm1 # 16-byte Reload546 movaps %xmm1, %xmm0547 movaps 48(%rsp), %xmm3 # 16-byte Reload548 shufps $0, %xmm3, %xmm0 # xmm0 = xmm0[0,0],xmm3[0,0]549 shufps $36, %xmm0, %xmm14 # xmm14 = xmm14[0,1],xmm0[2,0]550 movaps %xmm4, %xmm12551 shufps $17, %xmm15, %xmm12 # xmm12 = xmm12[1,0],xmm15[1,0]552 movaps %xmm3, %xmm2553 unpcklps %xmm1, %xmm2 # xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]554 shufps $226, %xmm2, %xmm12 # xmm12 = xmm12[2,0],xmm2[2,3]555 movaps %xmm15, %xmm7556 unpckhps %xmm4, %xmm7 # xmm7 = xmm7[2],xmm4[2],xmm7[3],xmm4[3]557 movaps %xmm1, %xmm2558 shufps $34, %xmm3, %xmm2 # xmm2 = xmm2[2,0],xmm3[2,0]559 shufps $36, %xmm2, %xmm7 # xmm7 = xmm7[0,1],xmm2[2,0]560 shufps $51, %xmm15, %xmm4 # xmm4 = xmm4[3,0],xmm15[3,0]561 unpckhps %xmm1, %xmm3 # xmm3 = xmm3[2],xmm1[2],xmm3[3],xmm1[3]562 shufps $226, %xmm3, %xmm4 # xmm4 = xmm4[2,0],xmm3[2,3]563 movaps %xmm4, (%rsp) # 16-byte Spill564 movaps -96(%r11), %xmm2565 movaps -80(%r11), %xmm1566 movaps %xmm1, %xmm4567 shufps $0, %xmm2, %xmm4 # xmm4 = xmm4[0,0],xmm2[0,0]568 movaps -112(%r11), %xmm10569 movaps -128(%r11), %xmm0570 movaps %xmm0, %xmm15571 unpcklps %xmm10, %xmm15 # xmm15 = xmm15[0],xmm10[0],xmm15[1],xmm10[1]572 shufps $36, %xmm4, %xmm15 # xmm15 = xmm15[0,1],xmm4[2,0]573 movaps %xmm2, %xmm4574 unpcklps %xmm1, %xmm4 # xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1]575 movaps %xmm10, %xmm6576 shufps $17, %xmm0, %xmm6 # xmm6 = xmm6[1,0],xmm0[1,0]577 shufps $226, %xmm4, %xmm6 # xmm6 = xmm6[2,0],xmm4[2,3]578 movaps %xmm1, %xmm3579 shufps $34, %xmm2, %xmm3 # xmm3 = xmm3[2,0],xmm2[2,0]580 movaps %xmm0, %xmm4581 unpckhps %xmm10, %xmm4 # xmm4 = xmm4[2],xmm10[2],xmm4[3],xmm10[3]582 shufps $36, %xmm3, %xmm4 # xmm4 = xmm4[0,1],xmm3[2,0]583 unpckhps %xmm1, %xmm2 # xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]584 shufps $51, %xmm0, %xmm10 # xmm10 = xmm10[3,0],xmm0[3,0]585 shufps $226, %xmm2, %xmm10 # xmm10 = xmm10[2,0],xmm2[2,3]586 movaps %xmm5, 192(%rsp) # 16-byte Spill587 mulps %xmm5, %xmm15588 addps %xmm14, %xmm15589 mulps %xmm5, %xmm6590 addps %xmm12, %xmm6591 mulps %xmm5, %xmm4592 addps %xmm7, %xmm4593 mulps %xmm5, %xmm10594 addps (%rsp), %xmm10 # 16-byte Folded Reload595 movaps %xmm4, %xmm0596 unpckhps %xmm10, %xmm0 # xmm0 = xmm0[2],xmm10[2],xmm0[3],xmm10[3]597 movaps %xmm6, %xmm1598 shufps $51, %xmm15, %xmm1 # xmm1 = xmm1[3,0],xmm15[3,0]599 shufps $226, %xmm0, %xmm1 # xmm1 = xmm1[2,0],xmm0[2,3]600 movaps %xmm1, 64(%rsp) # 16-byte Spill601 movaps %xmm10, %xmm0602 shufps $34, %xmm4, %xmm0 # xmm0 = xmm0[2,0],xmm4[2,0]603 movaps %xmm15, %xmm1604 unpckhps %xmm6, %xmm1 # xmm1 = xmm1[2],xmm6[2],xmm1[3],xmm6[3]605 shufps $36, %xmm0, %xmm1 # xmm1 = xmm1[0,1],xmm0[2,0]606 movaps %xmm1, 48(%rsp) # 16-byte Spill607 movaps %xmm4, %xmm0608 unpcklps %xmm10, %xmm0 # xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1]609 movaps %xmm6, %xmm1610 shufps $17, %xmm15, %xmm1 # xmm1 = xmm1[1,0],xmm15[1,0]611 shufps $226, %xmm0, %xmm1 # xmm1 = xmm1[2,0],xmm0[2,3]612 movaps %xmm1, (%rsp) # 16-byte Spill613 shufps $0, %xmm4, %xmm10 # xmm10 = xmm10[0,0],xmm4[0,0]614 unpcklps %xmm6, %xmm15 # xmm15 = xmm15[0],xmm6[0],xmm15[1],xmm6[1]615 shufps $36, %xmm10, %xmm15 # xmm15 = xmm15[0,1],xmm10[2,0]616 movaps %xmm11, %xmm10617 movaps 112(%rsp), %xmm14 # 16-byte Reload618 unpcklps %xmm14, %xmm10 # xmm10 = xmm10[0],xmm14[0],xmm10[1],xmm14[1]619 movaps 80(%rsp), %xmm2 # 16-byte Reload620 movaps %xmm2, %xmm0621 movaps 208(%rsp), %xmm3 # 16-byte Reload622 shufps $0, %xmm3, %xmm0 # xmm0 = xmm0[0,0],xmm3[0,0]623 shufps $36, %xmm0, %xmm10 # xmm10 = xmm10[0,1],xmm0[2,0]624 movaps %xmm14, %xmm12625 shufps $17, %xmm11, %xmm12 # xmm12 = xmm12[1,0],xmm11[1,0]626 movaps %xmm3, %xmm0627 unpcklps %xmm2, %xmm0 # xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]628 shufps $226, %xmm0, %xmm12 # xmm12 = xmm12[2,0],xmm0[2,3]629 movaps %xmm11, %xmm0630 unpckhps %xmm14, %xmm0 # xmm0 = xmm0[2],xmm14[2],xmm0[3],xmm14[3]631 movaps %xmm2, %xmm1632 shufps $34, %xmm3, %xmm1 # xmm1 = xmm1[2,0],xmm3[2,0]633 shufps $36, %xmm1, %xmm0 # xmm0 = xmm0[0,1],xmm1[2,0]634 shufps $51, %xmm11, %xmm14 # xmm14 = xmm14[3,0],xmm11[3,0]635 unpckhps %xmm2, %xmm3 # xmm3 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]636 shufps $226, %xmm3, %xmm14 # xmm14 = xmm14[2,0],xmm3[2,3]637 movaps -32(%r11), %xmm1638 movaps -16(%r11), %xmm2639 movaps %xmm2, %xmm3640 shufps $0, %xmm1, %xmm3 # xmm3 = xmm3[0,0],xmm1[0,0]641 movaps -48(%r11), %xmm4642 movaps -64(%r11), %xmm5643 movaps %xmm5, %xmm11644 unpcklps %xmm4, %xmm11 # xmm11 = xmm11[0],xmm4[0],xmm11[1],xmm4[1]645 shufps $36, %xmm3, %xmm11 # xmm11 = xmm11[0,1],xmm3[2,0]646 movaps %xmm1, %xmm3647 unpcklps %xmm2, %xmm3 # xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]648 movaps %xmm4, %xmm7649 shufps $17, %xmm5, %xmm7 # xmm7 = xmm7[1,0],xmm5[1,0]650 shufps $226, %xmm3, %xmm7 # xmm7 = xmm7[2,0],xmm3[2,3]651 movaps %xmm2, %xmm3652 shufps $34, %xmm1, %xmm3 # xmm3 = xmm3[2,0],xmm1[2,0]653 movaps %xmm5, %xmm6654 unpckhps %xmm4, %xmm6 # xmm6 = xmm6[2],xmm4[2],xmm6[3],xmm4[3]655 shufps $36, %xmm3, %xmm6 # xmm6 = xmm6[0,1],xmm3[2,0]656 unpckhps %xmm2, %xmm1 # xmm1 = xmm1[2],xmm2[2],xmm1[3],xmm2[3]657 shufps $51, %xmm5, %xmm4 # xmm4 = xmm4[3,0],xmm5[3,0]658 shufps $226, %xmm1, %xmm4 # xmm4 = xmm4[2,0],xmm1[2,3]659 movaps 192(%rsp), %xmm1 # 16-byte Reload660 mulps %xmm1, %xmm11661 addps %xmm10, %xmm11662 mulps %xmm1, %xmm7663 addps %xmm12, %xmm7664 mulps %xmm1, %xmm6665 addps %xmm0, %xmm6666 mulps %xmm1, %xmm4667 addps %xmm14, %xmm4668 movaps %xmm6, %xmm0669 unpckhps %xmm4, %xmm0 # xmm0 = xmm0[2],xmm4[2],xmm0[3],xmm4[3]670 movaps %xmm7, %xmm1671 shufps $51, %xmm11, %xmm1 # xmm1 = xmm1[3,0],xmm11[3,0]672 shufps $226, %xmm0, %xmm1 # xmm1 = xmm1[2,0],xmm0[2,3]673 movaps %xmm1, 80(%rsp) # 16-byte Spill674 movaps %xmm4, %xmm0675 shufps $34, %xmm6, %xmm0 # xmm0 = xmm0[2,0],xmm6[2,0]676 movaps %xmm11, %xmm12677 unpckhps %xmm7, %xmm12 # xmm12 = xmm12[2],xmm7[2],xmm12[3],xmm7[3]678 shufps $36, %xmm0, %xmm12 # xmm12 = xmm12[0,1],xmm0[2,0]679 movaps %xmm6, %xmm0680 unpcklps %xmm4, %xmm0 # xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]681 movaps %xmm7, %xmm1682 shufps $17, %xmm11, %xmm1 # xmm1 = xmm1[1,0],xmm11[1,0]683 shufps $226, %xmm0, %xmm1 # xmm1 = xmm1[2,0],xmm0[2,3]684 movaps %xmm1, 112(%rsp) # 16-byte Spill685 shufps $0, %xmm6, %xmm4 # xmm4 = xmm4[0,0],xmm6[0,0]686 unpcklps %xmm7, %xmm11 # xmm11 = xmm11[0],xmm7[0],xmm11[1],xmm7[1]687 shufps $36, %xmm4, %xmm11 # xmm11 = xmm11[0,1],xmm4[2,0]688 movaps %xmm9, %xmm10689 unpcklps %xmm13, %xmm10 # xmm10 = xmm10[0],xmm13[0],xmm10[1],xmm13[1]690 movaps 240(%rsp), %xmm2 # 16-byte Reload691 movaps %xmm2, %xmm0692 movaps 224(%rsp), %xmm3 # 16-byte Reload693 shufps $0, %xmm3, %xmm0 # xmm0 = xmm0[0,0],xmm3[0,0]694 shufps $36, %xmm0, %xmm10 # xmm10 = xmm10[0,1],xmm0[2,0]695 movaps %xmm13, %xmm14696 shufps $17, %xmm9, %xmm14 # xmm14 = xmm14[1,0],xmm9[1,0]697 movaps %xmm3, %xmm0698 unpcklps %xmm2, %xmm0 # xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]699 shufps $226, %xmm0, %xmm14 # xmm14 = xmm14[2,0],xmm0[2,3]700 movaps %xmm9, %xmm0701 unpckhps %xmm13, %xmm0 # xmm0 = xmm0[2],xmm13[2],xmm0[3],xmm13[3]702 movaps %xmm2, %xmm1703 shufps $34, %xmm3, %xmm1 # xmm1 = xmm1[2,0],xmm3[2,0]704 shufps $36, %xmm1, %xmm0 # xmm0 = xmm0[0,1],xmm1[2,0]705 shufps $51, %xmm9, %xmm13 # xmm13 = xmm13[3,0],xmm9[3,0]706 unpckhps %xmm2, %xmm3 # xmm3 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]707 shufps $226, %xmm3, %xmm13 # xmm13 = xmm13[2,0],xmm3[2,3]708 movaps 32(%r11), %xmm1709 movaps 48(%r11), %xmm2710 movaps %xmm2, %xmm3711 shufps $0, %xmm1, %xmm3 # xmm3 = xmm3[0,0],xmm1[0,0]712 movaps 16(%r11), %xmm4713 movaps (%r11), %xmm5714 movaps %xmm5, %xmm9715 unpcklps %xmm4, %xmm9 # xmm9 = xmm9[0],xmm4[0],xmm9[1],xmm4[1]716 shufps $36, %xmm3, %xmm9 # xmm9 = xmm9[0,1],xmm3[2,0]717 movaps %xmm1, %xmm3718 unpcklps %xmm2, %xmm3 # xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]719 movaps %xmm4, %xmm7720 shufps $17, %xmm5, %xmm7 # xmm7 = xmm7[1,0],xmm5[1,0]721 shufps $226, %xmm3, %xmm7 # xmm7 = xmm7[2,0],xmm3[2,3]722 movaps %xmm2, %xmm3723 shufps $34, %xmm1, %xmm3 # xmm3 = xmm3[2,0],xmm1[2,0]724 movaps %xmm5, %xmm6725 unpckhps %xmm4, %xmm6 # xmm6 = xmm6[2],xmm4[2],xmm6[3],xmm4[3]726 shufps $36, %xmm3, %xmm6 # xmm6 = xmm6[0,1],xmm3[2,0]727 unpckhps %xmm2, %xmm1 # xmm1 = xmm1[2],xmm2[2],xmm1[3],xmm2[3]728 shufps $51, %xmm5, %xmm4 # xmm4 = xmm4[3,0],xmm5[3,0]729 shufps $226, %xmm1, %xmm4 # xmm4 = xmm4[2,0],xmm1[2,3]730 movaps 192(%rsp), %xmm1 # 16-byte Reload731 mulps %xmm1, %xmm9732 addps %xmm10, %xmm9733 mulps %xmm1, %xmm7734 addps %xmm14, %xmm7735 mulps %xmm1, %xmm6736 addps %xmm0, %xmm6737 mulps %xmm1, %xmm4738 addps %xmm13, %xmm4739 movaps %xmm6, %xmm0740 unpckhps %xmm4, %xmm0 # xmm0 = xmm0[2],xmm4[2],xmm0[3],xmm4[3]741 movaps %xmm7, %xmm3742 shufps $51, %xmm9, %xmm3 # xmm3 = xmm3[3,0],xmm9[3,0]743 shufps $226, %xmm0, %xmm3 # xmm3 = xmm3[2,0],xmm0[2,3]744 movaps %xmm4, %xmm0745 shufps $34, %xmm6, %xmm0 # xmm0 = xmm0[2,0],xmm6[2,0]746 movaps %xmm9, %xmm2747 unpckhps %xmm7, %xmm2 # xmm2 = xmm2[2],xmm7[2],xmm2[3],xmm7[3]748 shufps $36, %xmm0, %xmm2 # xmm2 = xmm2[0,1],xmm0[2,0]749 movaps %xmm6, %xmm0750 unpcklps %xmm4, %xmm0 # xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]751 movaps %xmm7, %xmm13752 shufps $17, %xmm9, %xmm13 # xmm13 = xmm13[1,0],xmm9[1,0]753 shufps $226, %xmm0, %xmm13 # xmm13 = xmm13[2,0],xmm0[2,3]754 shufps $0, %xmm6, %xmm4 # xmm4 = xmm4[0,0],xmm6[0,0]755 movaps 256(%rsp), %xmm6 # 16-byte Reload756 movaps 272(%rsp), %xmm1 # 16-byte Reload757 unpcklps %xmm7, %xmm9 # xmm9 = xmm9[0],xmm7[0],xmm9[1],xmm7[1]758 shufps $36, %xmm4, %xmm9 # xmm9 = xmm9[0,1],xmm4[2,0]759 addq $6144, %r11 # imm = 0x1800760 addq $4, %r12761 jne .LBB5_7762# %bb.8: # %polly.loop_exit22763 # in Loop: Header=BB5_6 Depth=5764 movups %xmm8, (%r8)765 movaps 96(%rsp), %xmm0 # 16-byte Reload766 movups %xmm0, 16(%r8)767 movups %xmm6, 32(%r8)768 movups %xmm1, 48(%r8)769 movaps 64(%rsp), %xmm0 # 16-byte Reload770 movups %xmm0, 48(%rdx)771 movaps 48(%rsp), %xmm0 # 16-byte Reload772 movups %xmm0, 32(%rdx)773 movaps (%rsp), %xmm0 # 16-byte Reload774 movups %xmm0, 16(%rdx)775 movups %xmm15, (%rdx)776 movaps 80(%rsp), %xmm0 # 16-byte Reload777 movups %xmm0, 48(%r13)778 movaps 112(%rsp), %xmm0 # 16-byte Reload779 movups %xmm0, 16(%r13)780 movups %xmm11, (%r13)781 movups %xmm12, 32(%r13)782 movups %xmm3, 48(%r10)783 movups %xmm13, 16(%r10)784 movups %xmm9, (%r10)785 movups %xmm2, 32(%r10)786 addq $6144, %r9 # imm = 0x1800787 cmpq %rsi, %rax788 leaq 1(%rax), %rax789 jl .LBB5_6790# %bb.9: # %polly.loop_exit16791 # in Loop: Header=BB5_5 Depth=4792 movq 176(%rsp), %rax # 8-byte Reload793 addq $64, %rax794 addq $393216, %r14 # imm = 0x60000795 movq 184(%rsp), %r9 # 8-byte Reload796 addq $256, %r9 # imm = 0x100797 cmpq $1536, %rax # imm = 0x600798 movq 168(%rsp), %rdx # 8-byte Reload799 jb .LBB5_5800# %bb.10: # %polly.loop_exit10801 # in Loop: Header=BB5_4 Depth=3802 movq 144(%rsp), %rax # 8-byte Reload803 addq $64, %rax804 movq 152(%rsp), %rcx # 8-byte Reload805 incq %rcx806 movq 160(%rsp), %r14 # 8-byte Reload807 addq $256, %r14 # imm = 0x100808 cmpq $1536, %rax # imm = 0x600809 jb .LBB5_4810# %bb.11: # %polly.loop_exit4811 # in Loop: Header=BB5_3 Depth=2812 addq $64, %rdx813 addq $393216, 24(%rsp) # 8-byte Folded Spill814 # imm = 0x60000815 cmpq 136(%rsp), %rdx # 8-byte Folded Reload816 jle .LBB5_3817.LBB5_1: # %polly.par.setup818 # =>This Loop Header: Depth=1819 # Child Loop BB5_3 Depth 2820 # Child Loop BB5_4 Depth 3821 # Child Loop BB5_5 Depth 4822 # Child Loop BB5_6 Depth 5823 # Child Loop BB5_7 Depth 6824 leaq 40(%rsp), %rdi825 leaq 32(%rsp), %rsi826 callq GOMP_loop_runtime_next@PLT827 testb %al, %al828 jne .LBB5_2829# %bb.12: # %polly.par.exit830 callq GOMP_loop_end_nowait@PLT831 addq $296, %rsp # imm = 0x128832 .cfi_def_cfa_offset 56833 popq %rbx834 .cfi_def_cfa_offset 48835 popq %r12836 .cfi_def_cfa_offset 40837 popq %r13838 .cfi_def_cfa_offset 32839 popq %r14840 .cfi_def_cfa_offset 24841 popq %r15842 .cfi_def_cfa_offset 16843 popq %rbp844 .cfi_def_cfa_offset 8845 retq846.Lfunc_end5:847 .size main_polly_subfn_1, .Lfunc_end5-main_polly_subfn_1848 .cfi_endproc849 # -- End function850 .type A,@object # @A851 .comm A,9437184,16852 .type B,@object # @B853 .comm B,9437184,16854 .type .L.str,@object # @.str855 .section .rodata.str1.1,"aMS",@progbits,1856.L.str:857 .asciz "%lf "858 .size .L.str, 5859 860 .type C,@object # @C861 .comm C,9437184,16862 863 .ident "clang version 8.0.0 (trunk 342834) (llvm/trunk 342856)"864 .section ".note.GNU-stack","",@progbits865