polly/docs/experiments/matmul/matmul.polly.interchanged.s

   1         .text
   2         .file   "matmul.c"
   3         .section        .rodata.cst8,"aM",@progbits,8
   4         .p2align        3               # -- Begin function init_array
   5 .LCPI0_0:
   6         .quad   4602678819172646912     # double 0.5
   7         .text
   8         .globl  init_array
   9         .p2align        4, 0x90
  10         .type   init_array,@function
  11 init_array:                             # @init_array
  12         .cfi_startproc
  13 # %bb.0:                                # %entry
  14         pushq   %rbp
  15         .cfi_def_cfa_offset 16
  16         .cfi_offset %rbp, -16
  17         movq    %rsp, %rbp
  18         .cfi_def_cfa_register %rbp
  19         leaq    B(%rip), %rax
  20         leaq    A(%rip), %rcx
  21         xorl    %r8d, %r8d
  22         movsd   .LCPI0_0(%rip), %xmm0   # xmm0 = mem[0],zero
  23         xorl    %r9d, %r9d
  24         .p2align        4, 0x90
  25 .LBB0_1:                                # %polly.loop_header
  26                                         # =>This Loop Header: Depth=1
  27                                         #     Child Loop BB0_2 Depth 2
  28         movl    $1, %edi
  29         xorl    %edx, %edx
  30         .p2align        4, 0x90
  31 .LBB0_2:                                # %polly.loop_header1
  32                                         #   Parent Loop BB0_1 Depth=1
  33                                         # =>  This Inner Loop Header: Depth=2
  34         movl    %edx, %esi
  35         andl    $1022, %esi             # imm = 0x3FE
  36         orl     $1, %esi
  37         xorps   %xmm1, %xmm1
  38         cvtsi2sdl       %esi, %xmm1
  39         mulsd   %xmm0, %xmm1
  40         cvtsd2ss        %xmm1, %xmm1
  41         movss   %xmm1, -4(%rcx,%rdi,4)
  42         movss   %xmm1, -4(%rax,%rdi,4)
  43         leal    (%r9,%rdx), %esi
  44         andl    $1023, %esi             # imm = 0x3FF
  45         addl    $1, %esi
  46         xorps   %xmm1, %xmm1
  47         cvtsi2sdl       %esi, %xmm1
  48         mulsd   %xmm0, %xmm1
  49         cvtsd2ss        %xmm1, %xmm1
  50         movss   %xmm1, (%rcx,%rdi,4)
  51         movss   %xmm1, (%rax,%rdi,4)
  52         addq    $2, %rdi
  53         addl    %r8d, %edx
  54         cmpq    $1537, %rdi             # imm = 0x601
  55         jne     .LBB0_2
  56 # %bb.3:                                # %polly.loop_exit3
  57                                         #   in Loop: Header=BB0_1 Depth=1
  58         addq    $1, %r9
  59         addq    $6144, %rax             # imm = 0x1800
  60         addq    $6144, %rcx             # imm = 0x1800
  61         addl    $2, %r8d
  62         cmpq    $1536, %r9              # imm = 0x600
  63         jne     .LBB0_1
  64 # %bb.4:                                # %polly.exiting
  65         popq    %rbp
  66         .cfi_def_cfa %rsp, 8
  67         retq
  68 .Lfunc_end0:
  69         .size   init_array, .Lfunc_end0-init_array
  70         .cfi_endproc
  71                                         # -- End function
  72         .globl  print_array             # -- Begin function print_array
  73         .p2align        4, 0x90
  74         .type   print_array,@function
  75 print_array:                            # @print_array
  76         .cfi_startproc
  77 # %bb.0:                                # %entry
  78         pushq   %rbp
  79         .cfi_def_cfa_offset 16
  80         .cfi_offset %rbp, -16
  81         movq    %rsp, %rbp
  82         .cfi_def_cfa_register %rbp
  83         pushq   %r15
  84         pushq   %r14
  85         pushq   %r13
  86         pushq   %r12
  87         pushq   %rbx
  88         pushq   %rax
  89         .cfi_offset %rbx, -56
  90         .cfi_offset %r12, -48
  91         .cfi_offset %r13, -40
  92         .cfi_offset %r14, -32
  93         .cfi_offset %r15, -24
  94         leaq    C(%rip), %r13
  95         xorl    %eax, %eax
  96         movl    $3435973837, %r12d      # imm = 0xCCCCCCCD
  97         leaq    .L.str(%rip), %r14
  98         .p2align        4, 0x90
  99 .LBB1_1:                                # %for.cond1.preheader
 100                                         # =>This Loop Header: Depth=1
 101                                         #     Child Loop BB1_2 Depth 2
 102         movq    %rax, -48(%rbp)         # 8-byte Spill
 103         movq    stdout(%rip), %rsi
 104         xorl    %ebx, %ebx
 105         .p2align        4, 0x90
 106 .LBB1_2:                                # %for.body3
 107                                         #   Parent Loop BB1_1 Depth=1
 108                                         # =>  This Inner Loop Header: Depth=2
 109         movl    %ebx, %eax
 110         imulq   %r12, %rax
 111         shrq    $38, %rax
 112         leal    (%rax,%rax,4), %r15d
 113         shll    $4, %r15d
 114         addl    $79, %r15d
 115         movss   (%r13,%rbx,4), %xmm0    # xmm0 = mem[0],zero,zero,zero
 116         cvtss2sd        %xmm0, %xmm0
 117         movb    $1, %al
 118         movq    %rsi, %rdi
 119         movq    %r14, %rsi
 120         callq   fprintf
 121         cmpl    %ebx, %r15d
 122         jne     .LBB1_4
 123 # %bb.3:                                # %if.then
 124                                         #   in Loop: Header=BB1_2 Depth=2
 125         movq    stdout(%rip), %rsi
 126         movl    $10, %edi
 127         callq   fputc@PLT
 128 .LBB1_4:                                # %for.inc
 129                                         #   in Loop: Header=BB1_2 Depth=2
 130         addq    $1, %rbx
 131         movq    stdout(%rip), %rsi
 132         cmpq    $1536, %rbx             # imm = 0x600
 133         jne     .LBB1_2
 134 # %bb.5:                                # %for.end
 135                                         #   in Loop: Header=BB1_1 Depth=1
 136         movl    $10, %edi
 137         callq   fputc@PLT
 138         movq    -48(%rbp), %rax         # 8-byte Reload
 139         addq    $1, %rax
 140         addq    $6144, %r13             # imm = 0x1800
 141         cmpq    $1536, %rax             # imm = 0x600
 142         jne     .LBB1_1
 143 # %bb.6:                                # %for.end12
 144         addq    $8, %rsp
 145         popq    %rbx
 146         popq    %r12
 147         popq    %r13
 148         popq    %r14
 149         popq    %r15
 150         popq    %rbp
 151         .cfi_def_cfa %rsp, 8
 152         retq
 153 .Lfunc_end1:
 154         .size   print_array, .Lfunc_end1-print_array
 155         .cfi_endproc
 156                                         # -- End function
 157         .globl  main                    # -- Begin function main
 158         .p2align        4, 0x90
 159         .type   main,@function
 160 main:                                   # @main
 161         .cfi_startproc
 162 # %bb.0:                                # %entry
 163         pushq   %rbp
 164         .cfi_def_cfa_offset 16
 165         .cfi_offset %rbp, -16
 166         movq    %rsp, %rbp
 167         .cfi_def_cfa_register %rbp
 168         pushq   %r14
 169         pushq   %rbx
 170         .cfi_offset %rbx, -32
 171         .cfi_offset %r14, -24
 172         callq   init_array
 173         leaq    C(%rip), %rbx
 174         xorl    %r14d, %r14d
 175         xorl    %esi, %esi
 176         movl    $9437184, %edx          # imm = 0x900000
 177         movq    %rbx, %rdi
 178         callq   memset@PLT
 179         leaq    B(%rip), %rax
 180         leaq    A(%rip), %rcx
 181         .p2align        4, 0x90
 182 .LBB2_1:                                # %polly.loop_header8
 183                                         # =>This Loop Header: Depth=1
 184                                         #     Child Loop BB2_2 Depth 2
 185                                         #       Child Loop BB2_3 Depth 3
 186         movq    %rax, %rdx
 187         xorl    %esi, %esi
 188         .p2align        4, 0x90
 189 .LBB2_2:                                # %polly.loop_header14
 190                                         #   Parent Loop BB2_1 Depth=1
 191                                         # =>  This Loop Header: Depth=2
 192                                         #       Child Loop BB2_3 Depth 3
 193         leaq    (%r14,%r14,2), %rdi
 194         shlq    $11, %rdi
 195         addq    %rcx, %rdi
 196         movss   (%rdi,%rsi,4), %xmm0    # xmm0 = mem[0],zero,zero,zero
 197         shufps  $0, %xmm0, %xmm0        # xmm0 = xmm0[0,0,0,0]
 198         movl    $12, %edi
 199         .p2align        4, 0x90
 200 .LBB2_3:                                # %vector.body
 201                                         #   Parent Loop BB2_1 Depth=1
 202                                         #     Parent Loop BB2_2 Depth=2
 203                                         # =>    This Inner Loop Header: Depth=3
 204         movaps  -48(%rdx,%rdi,4), %xmm1
 205         mulps   %xmm0, %xmm1
 206         movaps  -32(%rdx,%rdi,4), %xmm2
 207         mulps   %xmm0, %xmm2
 208         addps   -48(%rbx,%rdi,4), %xmm1
 209         addps   -32(%rbx,%rdi,4), %xmm2
 210         movaps  %xmm1, -48(%rbx,%rdi,4)
 211         movaps  %xmm2, -32(%rbx,%rdi,4)
 212         movaps  -16(%rdx,%rdi,4), %xmm1
 213         mulps   %xmm0, %xmm1
 214         movaps  (%rdx,%rdi,4), %xmm2
 215         mulps   %xmm0, %xmm2
 216         addps   -16(%rbx,%rdi,4), %xmm1
 217         addps   (%rbx,%rdi,4), %xmm2
 218         movaps  %xmm1, -16(%rbx,%rdi,4)
 219         movaps  %xmm2, (%rbx,%rdi,4)
 220         addq    $16, %rdi
 221         cmpq    $1548, %rdi             # imm = 0x60C
 222         jne     .LBB2_3
 223 # %bb.4:                                # %polly.loop_exit22
 224                                         #   in Loop: Header=BB2_2 Depth=2
 225         addq    $1, %rsi
 226         addq    $6144, %rdx             # imm = 0x1800
 227         cmpq    $1536, %rsi             # imm = 0x600
 228         jne     .LBB2_2
 229 # %bb.5:                                # %polly.loop_exit16
 230                                         #   in Loop: Header=BB2_1 Depth=1
 231         addq    $1, %r14
 232         addq    $6144, %rbx             # imm = 0x1800
 233         cmpq    $1536, %r14             # imm = 0x600
 234         jne     .LBB2_1
 235 # %bb.6:                                # %polly.exiting
 236         xorl    %eax, %eax
 237         popq    %rbx
 238         popq    %r14
 239         popq    %rbp
 240         .cfi_def_cfa %rsp, 8
 241         retq
 242 .Lfunc_end2:
 243         .size   main, .Lfunc_end2-main
 244         .cfi_endproc
 245                                         # -- End function
 246         .type   A,@object               # @A
 247         .comm   A,9437184,16
 248         .type   B,@object               # @B
 249         .comm   B,9437184,16
 250         .type   .L.str,@object          # @.str
 251         .section        .rodata.str1.1,"aMS",@progbits,1
 252 .L.str:
 253         .asciz  "%lf "
 254         .size   .L.str, 5
 255
 256         .type   C,@object               # @C
 257         .comm   C,9437184,16
 258
 259         .ident  "clang version 8.0.0 (trunk 342834) (llvm/trunk 342856)"
 260         .section        ".note.GNU-stack","",@progbits